{"id":"7438564d-9dc8-45ff-8c5c-f3e62575c722","entityType":"agent","slug":"clawhub-whh110112-human-writing-skills","name":"Advanced Human Writing & AI Humanizer","canonicalUrl":"https://www.xpersona.co/agent/clawhub-whh110112-human-writing-skills","canonicalPath":"/agent/clawhub-whh110112-human-writing-skills","generatedAt":"2026-10-10T08:48:30.026Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":null},"description":"AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity Skill: Advanced Human Writing & AI Humanizer Owner: whh110112 Summary: AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity Tags: latest:0.15.3 Version history: v0.15.3 | 2026-09-29T12:02:20.699Z | user 中文 - 新增多语言改写保真审查，提示事实、归因、否定、范围等内容变化。 - 新增长篇作品按章节的整体结构审查，并完善覆盖回执。 - 增加多语言评测样例、文档及 CLI/MCP 测试。 - 保持按需加载，避免无关模块增加提示词负担。 English - Add multilingual rewrite-fidelity checks for chan","descriptionLabel":"Technical summary","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 1.8K downloads reported by the source. Last updated 10/10/2026.","installCommand":"clawhub skill install s17c1h8vj8rk22j14t28det89983nz1g:human-writing-skills","sourceUrl":"https://clawhub.ai/whh110112/human-writing-skills","homepage":"https://clawhub.ai/whh110112/skills/human-writing-skills","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/whh110112/human-writing-skills","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/whh110112/skills/human-writing-skills","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":65,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity Skill: Advanced Human Writing & AI Humanizer Owner: whh110112 Summ"},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":null},"protocols":[{"protocol":"OPENCLEW","label":"OpenClaw","status":"self-declared","notes":"Declared in the public agent profile."}],"capabilities":[],"verifiedCount":0,"selfDeclaredCount":1,"capabilityMatrix":{"rows":[{"key":"OPENCLEW","type":"protocol","support":"unknown","confidenceSource":"profile","notes":"Listed on profile"}],"flattenedTokens":"protocol:OPENCLEW|unknown|profile"}},"adoption":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":null},"stars":null,"forks":null,"downloads":1778,"packageName":null,"latestVersion":"0.15.3","tractionLabel":"1.8K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":null},"lastUpdatedAt":"2026-10-10T02:23:34.856Z","lastCrawledAt":"2026-10-10T02:23:34.856Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-11T02:23:34.856Z","lastVerifiedAt":null,"highlights":[{"version":"0.15.3","createdAt":"2026-09-29T12:02:20.699Z","changelog":"## 中文 - 新增多语言改写保真审查，提示事实、归因、否定、范围等内容变化。 - 新增长篇作品按章节的整体结构审查，并完善覆盖回执。 - 增加多语言评测样例、文档及 CLI/MCP 测试。 - 保持按需加载，避免无关模块增加提示词负担。 ## English - Add multilingual rewrite-fidelity checks for changes in claims, attribution, negation, and scope. - Add opt-in whole-book chapter review with coverage verification. - Expand multilingual evaluation fixtures, docs, CLI, MCP, and tests. Validation: 176 tests passed; GitHub CI passed on main.","fileCount":159,"zipByteSize":353242},{"version":"0.15.2","createdAt":"2026-09-11T09:51:42.264Z","changelog":"## 双向互动承接 修复小说中对白、无声动作和叙述者自身回应容易漏审的问题。分别追踪同一回合的言语与动作，确认受影响的人是否已经接收、拒绝、受阻或明确延迟回应；保留含蓄承接，避免每句强行添加动作。 - 小说生成与改写按需识别人际动作，复用 voice 模块。 - 自动流水线与长篇深度审查覆盖无声互动；跨块待回应项目交由统稿确认。 - 更新中英文 README、Skill 入口与 SkillHub 中文概述。 - 163 项自动测试通过，保留原有提示词预算；独立模型试审覆盖缺失回应、隐含完成、拒绝和跨块不确定性。 ## Bidirectional Interaction Uptake Track each affected participant and separate simultaneous speech and action obligations. Preserve implied completion and evidenced delay; do not assume acceptance, contact, or following from an initiator's action alone. On-demand generation, automatic pipelines, and deep chunk audits now route directed wordless interactions. Reconciliation checks pending responses across chunk boundaries. Serious documents remain outside automatic fiction-interaction routing. Validated with 163 automated tests and a focused independent model review. Contextual response assessment is model-executed; these checks do not guarantee detection of every omission.","fileCount":151,"zipByteSize":333927},{"version":"0.15.1","createdAt":"2026-09-03T04:20:57.160Z","changelog":"Exclude Cursor editor-rule files from the SkillHub upload bundle. The GitHub repository retains the Cursor template; SkillHub receives only supported runtime and documentation files.","fileCount":150,"zipByteSize":328809},{"version":"0.14.2","createdAt":"2026-09-02T13:09:21.041Z","changelog":"## Narrative repetition and exposition audit - Add a narrative-only audit for exact echoes, repeated action choreography, inventory drift, and redundant explanatory narration. - Add the repetition isolated audit profile and a signal-gated long-form pipeline pass. - Keep quick humanization and serious-document workflows free of this narrative-only cost. - Add deterministic lint findings REP001, NAT005, and NAT006, with allow-list support. - Update English and Chinese guidance and packaging coverage.","fileCount":142,"zipByteSize":311095},{"version":"0.14.1","createdAt":"2026-09-01T03:23:18.881Z","changelog":"**Full Changelog**: https://github.com/whh110112/human-writing-skills/compare/v0.14.0...v0.14.1","fileCount":141,"zipByteSize":305064},{"version":"0.14.0","createdAt":"2026-09-01T03:20:23.559Z","changelog":"**Full Changelog**: https://github.com/whh110112/human-writing-skills/compare/v0.13.0...v0.14.0","fileCount":141,"zipByteSize":304872},{"version":"0.13.0","createdAt":"2026-08-30T14:11:57.203Z","changelog":"## Dialogue naturalness\\n\\n- Added dialogue-performance-audit for listener uptake, purposeful physical beats, and earned dialogue landings in fiction and webnovels.\\n- Explicit interaction requests such as reunion, testing, reconciliation, confrontation, negotiation, and dialogue activate the detailed modules on demand; narration-only and serious documents do not.\\n- Automatic pipeline review now recognizes a real two-turn exchange; default full review remains lightweight.\\n\\n## Reliability\\n\\n- The full ending module remains explicitly gated; ai-trace and default full review keep the lightweight END001 preflight.\\n- Local skill resources are covered by tests before packaging.\\n- Chinese and English documentation plus SkillHub discovery text now cover dialogue audit, dialogue action audit, and natural character interaction.\\n\\n## Verification\\n\\n- python -m unittest discover -s tests -v (129 tests)\\n- SKILL metadata validation\\n- SkillHub package and wheel content checks","fileCount":124,"zipByteSize":267720},{"version":"0.12.1","createdAt":"2026-08-24T05:49:05.130Z","changelog":"## Token-control fix - Keeps the full earned-ending module behind explicit \u0007udit --profile ending selection. - Retains deterministic END001 narrative-exit preflight without adding prompt tokens. - Keeps ull, \u0007i-trace, and automatic pipeline stacks from loading the specialist ending module implicitly. - Aligns the English and Chinese README with the actual activation contract. ## Verification - 129 tests passed locally and on Python 3.9, 3.10, and 3.12. - Codex skill validation passed. This patch supersedes v0.12.0.","fileCount":123,"zipByteSize":264475}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s17c1h8vj8rk22j14t28det89983nz1g:human-writing-skills","setupComplexity":"low","setupSteps":["Setup complexity is classified as HIGH. You must provision dedicated cloud infrastructure or an isolated VM. Do not run this directly on your local workstation.","Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data."],"contract":{"contractStatus":"missing","authModes":[],"requires":[],"forbidden":[],"supportsMcp":false,"supportsA2a":false,"supportsStreaming":false,"inputSchemaRef":null,"outputSchemaRef":null,"dataRegion":null,"contractUpdatedAt":null,"sourceUpdatedAt":null,"freshnessSeconds":null},"invocationGuide":{"preferredApi":{"snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-10T08:48:30.022Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-whh110112-human-writing-skills/trust"}},"reliability":{"evidence":{"source":"runtime-metrics","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No trust, reliability, or runtime telemetry is available."},"trust":{"status":"unavailable","handshakeStatus":"UNKNOWN","verificationFreshnessHours":null,"reputationScore":null,"p95LatencyMs":null,"successRate30d":null,"fallbackRate":null,"attempts30d":null,"trustUpdatedAt":null,"trustConfidence":"unknown","sourceUpdatedAt":null,"freshnessSeconds":null},"decisionGuardrails":{"doNotUseIf":["Contract metadata is missing or unavailable for deterministic execution."],"safeUseWhen":[],"riskFlags":["missing_or_unavailable_contract","trust_data_unavailable","schema_references_missing"],"operationalConfidence":"low"},"executionMetrics":{"observedLatencyMsP50":null,"observedLatencyMsP95":null,"estimatedCostUsd":null,"uptime30d":null,"rateLimitRpm":null,"rateLimitBurst":null,"lastVerifiedAt":null,"verificationSource":null},"runtimeMetrics":{"successRate":null,"avgLatencyMs":null,"avgCostUsd":null,"hallucinationRate":null,"retryRate":null,"disputeRate":null,"p50Latency":null,"p95Latency":null,"lastUpdated":null}},"benchmarks":{"evidence":{"source":"no-benchmark-data","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No benchmark suites or observed failure patterns are available."},"suites":[],"failurePatterns":[]},"artifacts":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"high","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":null},"readme":"Skill: Advanced Human Writing & AI Humanizer\n\nOwner: whh110112\n\nSummary: AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity\n\nTags: latest:0.15.3\n\nVersion history:\n\nv0.15.3 | 2026-09-29T12:02:20.699Z | user\n\n## 中文\n- 新增多语言改写保真审查，提示事实、归因、否定、范围等内容变化。\n- 新增长篇作品按章节的整体结构审查，并完善覆盖回执。\n- 增加多语言评测样例、文档及 CLI/MCP 测试。\n- 保持按需加载，避免无关模块增加提示词负担。\n\n## English\n- Add multilingual rewrite-fidelity checks for changes in claims, attribution, negation, and scope.\n- Add opt-in whole-book chapter review with coverage verification.\n- Expand multilingual evaluation fixtures, docs, CLI, MCP, and tests.\n\nValidation: 176 tests passed; GitHub CI passed on main.\n\nv0.15.2 | 2026-09-11T09:51:42.264Z | user\n\n## 双向互动承接\n\n修复小说中对白、无声动作和叙述者自身回应容易漏审的问题。分别追踪同一回合的言语与动作，确认受影响的人是否已经接收、拒绝、受阻或明确延迟回应；保留含蓄承接，避免每句强行添加动作。\n\n- 小说生成与改写按需识别人际动作，复用 voice 模块。\n- 自动流水线与长篇深度审查覆盖无声互动；跨块待回应项目交由统稿确认。\n- 更新中英文 README、Skill 入口与 SkillHub 中文概述。\n- 163 项自动测试通过，保留原有提示词预算；独立模型试审覆盖缺失回应、隐含完成、拒绝和跨块不确定性。\n\n## Bidirectional Interaction Uptake\n\nTrack each affected participant and separate simultaneous speech and action obligations. Preserve implied completion and evidenced delay; do not assume acceptance, contact, or following from an initiator's action alone.\n\nOn-demand generation, automatic pipelines, and deep chunk audits now route directed wordless interactions. Reconciliation checks pending responses across chunk boundaries. Serious documents remain outside automatic fiction-interaction routing.\n\nValidated with 163 automated tests and a focused independent model review. Contextual response assessment is model-executed; these checks do not guarantee detection of every omission.\n\nv0.15.1 | 2026-09-03T04:20:57.160Z | user\n\nExclude Cursor editor-rule files from the SkillHub upload bundle. The GitHub repository retains the Cursor template; SkillHub receives only supported runtime and documentation files.\n\nv0.14.2 | 2026-09-02T13:09:21.041Z | user\n\n## Narrative repetition and exposition audit\n\n- Add a narrative-only audit for exact echoes, repeated action choreography, inventory drift, and redundant explanatory narration.\n- Add the repetition isolated audit profile and a signal-gated long-form pipeline pass.\n- Keep quick humanization and serious-document workflows free of this narrative-only cost.\n- Add deterministic lint findings REP001, NAT005, and NAT006, with allow-list support.\n- Update English and Chinese guidance and packaging coverage.\n\nv0.14.1 | 2026-09-01T03:23:18.881Z | user\n\n**Full Changelog**: https://github.com/whh110112/human-writing-skills/compare/v0.14.0...v0.14.1\n\nv0.14.0 | 2026-09-01T03:20:23.559Z | user\n\n**Full Changelog**: https://github.com/whh110112/human-writing-skills/compare/v0.13.0...v0.14.0\n\nv0.13.0 | 2026-08-30T14:11:57.203Z | user\n\n## Dialogue naturalness\\n\\n- Added dialogue-performance-audit for listener uptake, purposeful physical beats, and earned dialogue landings in fiction and webnovels.\\n- Explicit interaction requests such as reunion, testing, reconciliation, confrontation, negotiation, and dialogue activate the detailed modules on demand; narration-only and serious documents do not.\\n- Automatic pipeline review now recognizes a real two-turn exchange; default full review remains lightweight.\\n\\n## Reliability\\n\\n- The full ending module remains explicitly gated; ai-trace and default full review keep the lightweight END001 preflight.\\n- Local skill resources are covered by tests before packaging.\\n- Chinese and English documentation plus SkillHub discovery text now cover dialogue audit, dialogue action audit, and natural character interaction.\\n\\n## Verification\\n\\n- python -m unittest discover -s tests -v (129 tests)\\n- SKILL metadata validation\\n- SkillHub package and wheel content checks\n\nv0.12.1 | 2026-08-24T05:49:05.130Z | user\n\n## Token-control fix\n\n- Keeps the full earned-ending module behind explicit \u0007udit --profile ending selection.\n- Retains deterministic END001 narrative-exit preflight without adding prompt tokens.\n- Keeps \full, \u0007i-trace, and automatic pipeline stacks from loading the specialist ending module implicitly.\n- Aligns the English and Chinese README with the actual activation contract.\n\n## Verification\n\n- 129 tests passed locally and on Python 3.9, 3.10, and 3.12.\n- Codex skill validation passed.\n\nThis patch supersedes v0.12.0.\n\nv0.12.0 | 2026-08-24T05:26:52.093Z | user\n\n## Highlights\n\n- Adds an earned-ending audit for Chinese and English AI-style reflective bookends, scenic dissolves, false closure, stock kickers, and meaningless uplift after the last real scene change.\n- Adds deterministic END001 evidence spans for narrative exits, with deletion-first repair and allowlisting.\n- Uses separate ending contracts for fiction/webnovels, hard news/features, academic/technical writing, and official documents.\n- Adds executable chunk-audit for cross-month and cross-model style unification, character consistency, dialogue voice, terminology, and continuity.\n- Keeps token use controlled: the full ending module is explicit/on-demand; automatic pipelines avoid duplicating it when ai-trace already covers the pass.\n- Improves English GitHub/ClawHub discovery and Chinese SkillHub discovery for AI story ending, scene ending audit, AI式结尾, 生硬结尾审查, 伪收束, 无意义升华, 长篇审查, and 文风统一.\n\n## Verification\n\n- 129 tests passed across Python 3.9, 3.10, and 3.12.\n- Codex skill validation passed.\n- Wheel and SkillHub publication packages were built and inspected.\n\nThe project improves editing quality and does not claim detector evasion or authorship classification.\n\nv0.10.7 | 2026-08-13T11:00:40.324Z | user\n\nFixes automated review findings for the narrative naturalness audit.\\n\\n- Keeps narrative-only audit instructions out of serious-document full audits.\\n- Requires affect nouns after Chinese quantifiers before counting vague-affect recurrence.\\n- Treats adjacent verbal dialogue turns as valid responses.\\n- Adds regression tests and keeps package metadata at 0.10.7.\n\nv0.10.6 | 2026-08-13T08:30:46.954Z | user\n\n## v0.10.6\n\n### Added\n- Added the on-demand `narrative-naturalness-audit` module for deep or explicit AI-trace review of fiction, webnovels, and narrative self-media.\n- Added deterministic `NAT001`-`NAT004` findings for repeated scene-entry recipes, vague-affect recurrence, abstract paragraph closures, and clustered pressure-bearing dialogue without visible uptake.\n\n### Guardrails\n- Quick humanization remains unchanged and does not load the new module.\n- Serious academic, legal, technical, formal, and news writing remains isolated from narrative naturalness rules.\n- Deep humanize prompt stays under the existing 30,000-character budget.\n- The audit improves prose quality and continuity; it does not promise detector evasion or authorship classification results.\n\n### Validation\n- 113 tests passed locally.\n- GitHub Actions passed on Python 3.9, 3.10, and 3.12.\n\nv0.10.5 | 2026-08-13T05:15:22.329Z | user\n\n## 中文\n\n- SkillHub 概述改为完整中文，面向中文用户说明去AI味、AI文本润色、小说续写、长文一致性、人物对白、空间物理与严肃文本保护等能力。\n- SkillHub 中文概述增加“高级 AI 写作工具”“去AI写作”“消除AI腔”“小说润色”“网络小说创作”等自然检索入口。\n- SkillHub 与 ClawHub 使用独立市场概述：中文市场不再显示英文主体，英文市场不受中文概述替换影响。\n- ClawHub 强化 AI text humanizer、de-AI writing、natural rewriting、fiction editing、novel writing assistant 与 story continuity 等英文检索词。\n\n## English\n\n- Adds a dedicated Chinese overview for SkillHub while keeping ClawHub English-first.\n- Improves English discovery for AI text humanizer, de-AI writing, natural rewriting, fiction editing, novel assistance, and story continuity.\n\nv0.10.4 | 2026-08-13T04:57:52.817Z | user\n\n- 修复 SkillHub 发布包包含平台不允许的 Git 仓库管理文件问题。\n- 保留增强版去 AI 写作 Skill、高级 AI 写作工具、去AI味、AI文本润色、小说润色等中文检索元数据。\n- SkillHub 与 ClawHub 自动发布均已配置。\n- 109 项测试与 SkillHub 发布预检通过。\n\nv0.10.3 | 2026-08-13T04:33:04.462Z | user\n\n- 保留 v0.10.2 的 SkillHub 中文搜索说明与 ClawHub 检索优化。\n- 修复 SkillHub 打包脚本在 Python 3.9 下的兼容性问题。\n- 已通过 109 项测试与 SkillHub 发布预检。\n\nKeeps the marketplace discovery improvements from v0.10.2 and fixes SkillHub package generation on Python 3.9.\n\nv0.10.2 | 2026-08-13T04:30:45.245Z | user\n\n## 中文\n\n- 优化 SkillHub 中文标题、摘要、说明与标签，覆盖“增强版去 AI 写作 Skill”“去AI味”“高级 AI 写作工具”“AI文本润色”“小说润色”等检索词。\n- 优化 ClawHub 英文摘要和主题词，增强 `AI humanizer`、`de-AI writing`、`human writing`、小说编辑与长篇连续性等发现入口。\n- 中英文 README 首屏补充项目定位，并继续强调项目包含可执行 Python CLI、确定性 lint/fix/verify 与按需加载模块。\n- SkillHub 未配置密钥时发布流程会安全跳过，不再误报整个 Release 失败。\n\n## English\n\n- Improved SkillHub Chinese title, summary, description, and search tags.\n- Refined ClawHub summary and topics for AI humanizer, de-AI writing, fiction editing, and story continuity discovery.\n- Clarified executable CLI and on-demand skill architecture in both READMEs.\n- Made SkillHub publication skip cleanly when its repository secret is not configured.\n\nv0.10.0 | 2026-08-13T03:40:51.540Z | user\n\n## 中文\n\n- 新增 `speech-register-continuity`：按人物经历、共同语言、听众、敬语、语气词和变调依据生成与审核对白，防止方言、外语和人物口吻串台。\n- 新增 `capability-state-audit`：区分永久战力/能力与临时伤势、装备、资源、冷却、克制和权限，要求变化与意外结果具备过程门和代价。\n- 新增 `formal-document` 基础文体，区分小说口语、公文、新闻与论文的语气密度，并为公文启用严肃文本保护。\n- 长篇上下文采用“权威账本 -> 最新状态 -> 近期章节 -> 相关旧片段 -> 不确定推断”的证据顺序，检索片段不能覆盖后续确认状态。\n- 新模块继续按需加载：`register` 只在对白与语言/身份证据同时存在时激活；`capability` 自动审查需要上下文；普通文戏不承担额外 Token。\n- 延长 ClawHub 最终一致性等待窗口，减少实际发布成功但 Action 误报失败的情况。\n\n## English\n\n- Adds evidence-gated speech register and language-identity continuity for dialects, honorifics, particles, address forms, and code-switching.\n- Adds capability-state continuity for power, skills, authority, equipment, injuries, resources, cooldowns, counters, costs, and earned transitions.\n- Adds a formal-document base style and separates conversational fiction from official, news, and academic register.\n- Defines a layered long-form context order so retrieved older passages cannot overwrite newer canonical state.\n- Keeps the new modules demand-loaded to preserve Token budgets.\n\n## Verification\n\n- 107 tests passed\n- Skill structure validation passed\n- PEP 517 wheel build passed\n- Compact review prompt: 21,629 characters\n- Quick humanize prompt: 12,613 characters\n- Deep humanize prompt: 28,358 characters\n\nv0.9.2 | 2026-08-11T07:39:48.284Z | user\n\nAdds a token-aware `humanize` shortcut with quick and deep modes, explicit voice and ambiguity preservation, an opt-in before/after example library, bilingual discovery metadata, and isolated preservation audits. Quick mode remains below the prompt-budget guard; examples and high-cost review passes stay opt-in.\n\nv0.9.1 | 2026-08-11T03:40:06.237Z | user\n\nAdds natural multilingual scene transitions, narrative mini-heading and time-card audits, broader language statistics, and multilingual discovery metadata.\n\nv0.9.0 | 2026-08-11T03:05:48.363Z | user\n\nAdds rewrite fidelity, genre-aware surface-pattern audits, comparison-ladder escalation, optional style statistics, conservative fixes, and bilingual documentation.\n\nv0.8.1 | 2026-08-06T02:19:24.659Z | user\n\n## 中文\n\n- 新增双向公式句式审查：识别“不是 X，是 Y”和“是 X，不是 Y”，并按复现密度升级。\n- 新增连续“比”比较阶梯检查，保留事实纠错、严肃比较和符合人物口吻的反驳。\n- 新增高置信漏字症状检测，以及按谓词槽位、并列结构和指代关系运行的完整终校流程。\n- 普通生成仅加载轻量句法完整性检查；完整漏字审查仍按需进入 `proofread` 或多阶段流水线。\n- 完善中英文 README、规则文档和测试覆盖。\n\n## English\n\n- Detects bidirectional formulaic contrast frames and escalates repeated density.\n- Audits chained Chinese comparison ladders while preserving necessary factual and serious comparisons.\n- Adds high-confidence omission linting plus a full sentence-skeleton proofreading pass.\n- Keeps ordinary generation lightweight; full omission review remains on demand through `proofread` or the pipeline.\n- Updates bilingual documentation and regression coverage.\n\n## Verification\n\n- 73 unit tests passed\n- Skill structure validation passed\n- PEP 517 wheel build passed\n\nv0.8.0 | 2026-08-05T17:19:45.059Z | user\n\nAdd evidence-driven deep writing audits, world and process checks, attention-budget review, chapter-pattern analysis, and improved dialogue/subtext consistency.\n\nv0.1.0 | 2026-07-12T10:31:40.446Z | auto\n\nInitial release of human-writing-skills.\n\n- Supports writing, rewriting, and auditing natural, genre-aware long-form prose.\n- Ensures physical, relationship, and fact continuity; enables protected fact verification and deterministic pattern linting.\n- Workflow includes modular style selection and explicit handling of style calibration from user-supplied samples.\n- Provides commands for building, auditing, linting, and verifying manuscripts across fiction, webnovels, essays, news, self-media, and academic prose.\n- Comprehensive user and developer documentation referenced in README and docs.\n\nArchive index:\n\nArchive v0.15.3: 159 files, 353242 bytes\n\nFiles: agents/openai.yaml (435b), CONTRIBUTING.md (1300b), docs/agent-orchestration.md (5277b), docs/agent-orchestration.zh-CN.md (4819b), docs/audit-pipeline.md (6033b), docs/audit-pipeline.zh-CN.md (6504b), docs/chatbox.md (7460b), docs/chatbox.zh-CN.md (7173b), docs/editing-tools.md (4202b), docs/editing-tools.zh-CN.md (3869b), docs/evaluation.md (2537b), docs/evaluation.zh-CN.md (1897b), docs/forum-complaint-research.md (4820b), docs/forum-complaint-research.zh-CN.md (4977b), docs/ledger-extraction.md (1396b), docs/ledger-extraction.zh-CN.md (1290b), docs/long-form-consistency.md (6250b), docs/long-form-consistency.zh-CN.md (7203b), docs/number-sense.md (2079b), docs/number-sense.zh-CN.md (2336b), docs/pattern-linter.md (5135b), docs/pattern-linter.zh-CN.md (4942b), docs/physical-continuity.md (3075b), docs/physical-continuity.zh-CN.md (3753b), docs/protected-content.md (2804b), docs/protected-content.zh-CN.md (1984b), docs/reference-style.md (1859b), docs/reference-style.zh-CN.md (1802b), docs/relationship-stance-continuity.md (3247b), docs/relationship-stance-continuity.zh-CN.md (3333b), examples/article-brief.md (622b), examples/capacity-conflict-draft.zh-CN.md (628b), examples/capacity-ledger-template.md (1103b), examples/chatbox-ledger-template.md (3620b), examples/deep-fiction-task.md (1465b), examples/false-precision-draft.zh-CN.md (706b), examples/problem-car-scene-draft.md (835b), examples/problem-car-scene-draft.zh-CN.md (683b), examples/reference-style-draft.zh-CN.md (184b), examples/reference-style-source.zh-CN.md (371b), examples/relationship-stance-ledger.zh-CN.md (1992b), examples/story-ledger.md (4053b), examples/vehicle-scene-ledger.md (3995b), humanwriting/__init__.py (127b), humanwriting/audit_queue.py (9005b), humanwriting/cli.py (30889b), humanwriting/compiler.py (36582b), humanwriting/config.py (3848b), humanwriting/detection.py (18713b), humanwriting/evaluation.py (4974b), humanwriting/fidelity.py (7169b), humanwriting/fixer.py (4556b), humanwriting/ledger.py (2446b), humanwriting/linter.py (61773b), humanwriting/longform.py (36285b), humanwriting/mcp_server.py (26062b), humanwriting/original.py (1341b), humanwriting/pipeline.py (13032b), humanwriting/precommit.py (1105b), humanwriting/protection.py (12274b), humanwriting/reference.py (3700b), humanwriting/skills.py (2090b), humanwriting/source.py (1814b), humanwriting/statistics.py (9377b), integrations/mcp/mcp.json.example (202b), integrations/project-context/AGENTS.md.example (741b), LICENSE (1107b), marketplaces/skillhub-overview.zh-CN.md (5791b), plugins/deepseek-harness/cordis.patch.yml (351b), plugins/deepseek-harness/index.js (840b), plugins/deepseek-harness/LICENSE (1092b), plugins/deepseek-harness/package.json (1158b), plugins/deepseek-harness/README.md (1278b), plugins/deepseek-harness/README.zh-CN.md (1005b), pyproject.toml (1378b), README.es.md (5497b), README.fr.md (5536b), README.md (47555b), README.pt-BR.md (5186b), README.zh-CN.md (47552b)\n\nFile v0.15.3:SKILL.md\n\n---\nname: human-writing-skills\ndescription: Advanced multilingual AI humanizer for natural rewriting, fiction editing, long-form audit and continuity, verified chunked agent review, translationese review, character consistency, opt-in whole-book story architecture review, and semantic rewrite-fidelity triage. Humanize AI text, remove robotic tone, edit fiction and novels, continue webnovel chapters, proofread writing, and audit story continuity, character voice, dialogue register and performance, scene geography, relationships, numbers, citations, source meaning, and translated-text fidelity. Use for AI writing cleanup, supplied-sample style matching, long-context fiction, essays, news, official, academic, legal, and technical prose. Trigger on humanize AI text, de-AI writing, natural rewriting, novel writing assistant, story consistency checker, scene ending audit, reflective ending, chunked audit, long-form agent audit, whole-book review, translationese audit, style consistency review, character consistency audit, dialogue audit, dialogue action audit, 增强版去 AI 写作 Skill、高级 AI 写作工具、去AI味、去AI写作、消除AI腔、AI人性化改写、AI文本润色、AI文章润色、小说润色、小说续写、AI式结尾、生硬结尾审查、无意义升华、长文一致性、长篇审查、分块审查、全书审查、文风统一、统一文风、人物设定统一、人物一致性审查、跨章一致性、小说审查、报告审查、人物口吻、人物对白审查、对话生硬、对白动作、方言语域、翻译腔、翻译审查、战力设定、场景空间审查、语义保真审查.\n---\n\n# Advanced Human Writing & AI Humanizer\n\nHumanize AI-shaped text, write or continue genre-aware prose, and audit long-form\ncontinuity without flattening a specific voice. Use the smallest set of modules\nthat covers the task. Keep project facts, prior chapters, rewrite originals, and\ncontinuity ledgers separate from optional style references.\n\n## Capability Layers\n\n- The `humanwriting/` Python package is executable. Its CLI deterministically\n  compiles prompts, locates recurring text patterns, calculates diagnostics,\n  previews conservative fixes, checks protected content, and writes staged audits.\n- The Markdown files under `skills/` are model-executed writing and editorial\n  modules. They are selected by the compiler and are intentionally not pretending\n  to be deterministic NLP algorithms.\n- A normal installation includes both layers. Verify the executable layer with\n  `human-writing-skills list --kind module` and run a real draft through `lint`,\n  `fix`, `verify`, or `pipeline` rather than judging the package from `SKILL.md` alone.\n- `human-writing-mcp` is an optional project-local coordination layer. It gives\n  separate agents bounded long-form assignments, stores receipts, and refuses\n  reconciliation until coverage verification passes. It does not call a model.\n\n## Quick Humanize Route\n\n- For a supplied draft, use `humanize --mode quick` for surface patterns,\n  rewrite fidelity, and voice/ambiguity preservation.\n- Use `--mode deep` only when structural repetition, cliches, paragraph stagnation,\n  or broad editorial reconstruction needs a separate pass.\n- Load `humanize-examples` only after an explicit request or `--with-examples`.\n- Do not activate rewrite-preservation modules for unrelated new drafting.\n\n## Workflow\n\n1. Select one base style from `skills/`: `fiction`, `webnovel`, `argumentative`,\n   `news-report`, `formal-document`, `self-media`, or `academic-paper`.\n2. Read only the relevant modules. Add continuity, spatial, relationship, number,\n   dialogue, dialogue-performance, register, capability, world, process, salience, recurrence, source, rhythm, preservation,\n   or AI-trace modules when the text actually needs them.\n   `narrative-naturalness-audit` is reserved for deep or explicit AI-trace review of\n   narrative prose; it is not loaded for ordinary quick humanization or serious documents.\n   For fiction with spoken or wordless interpersonal exchanges, use\n   `dialogue-voice-audit` and `dialogue-performance-audit` (CLI: `--profile voice`).\n   Check both participants and separate speech from simultaneous action; establish\n   reception, refusal, or an evidenced pending state before assuming completion.\n   During drafting, reconsider this route when an interaction enters the scene even\n   if the original request only said to continue a chapter.\n3. Treat user facts and `--context` as authoritative. Never borrow facts from a\n   style sample. When `--original` is supplied for a rewrite, activate\n   `rewrite-fidelity` and preserve meaning without preserving awkward wording.\n   In fiction and webnovels, do not let time/place mini-headings replace scene bridges.\n4. Activate `reference-style-alignment` only when the user supplies reference\n   material, gives an explicit style direction, or directly asks to match a style.\n5. Treat `--source` as factual evidence only. Activate `source-grounding` only for\n   serious academic, formal, news, legal, or technical work with explicit source files.\n6. For important revisions, run deterministic `lint`, then independent audit\n   profiles, then `verify` protected literals against the source. For consequential\n   rewrites use `verify-fidelity` too: changed claims require review even if every\n   number and citation is unchanged. Run `stats` only\n   when distributional diagnostics help; use `fix` as a preview before writing.\n7. Keep `voice`, `serial`, `world`, `process`, `momentum`, `salience`, `recurrence`,\n   `ending`, `texture`, `fidelity`, `preservation`, examples, and `sources` separate from the default audit. Activate them explicitly\n   or through `pipeline --auto`. `serial` requires context, `recurrence` requires at\n   least three chapters, `fidelity` requires `--original`, `preservation` requires\n   both explicit selection and `--original`, examples require an explicit request,\n   and `sources` requires both a serious document and `--source`.\n8. For a book-length manuscript or large report, use `chunk-audit` instead of placing\n   the whole draft in one prompt. Supply `--outline` or `--context` for authoritative\n   character or report rules. Use `--agent-mode deep` only when every block needs a\n   visible coverage receipt, then run `verify-chunk-audit` before reconciliation. Use\n   `--translationese` only for an explicitly translated or localized work. Read\n   `docs/long-form-consistency.md` only for this workflow.\n   Add `--book-level` only for an explicitly requested whole-book fiction/webnovel\n   review with at least three identifiable chapters; run it after chunk reports.\n9. When multiple agents or sessions share a long-form project, use\n   `human-writing-mcp` rather than copying the entire manuscript into each chat.\n   Agents must claim a task, submit a complete receipt, and pass coverage\n   verification before one final reconciliation task is released. Read\n   `docs/agent-orchestration.md` only when an MCP-capable host is available.\n\n## Commands\n\n```powershell\nhuman-writing-skills build --style fiction --context ledger.md --task \"Continue the scene.\"\nhuman-writing-skills humanize --draft chapter.md --style fiction --mode quick\nhuman-writing-skills humanize --draft article.md --style self-media --mode deep --with-examples\nhuman-writing-skills build --style fiction --reference sample.md --task \"Match the sample's restrained rhythm.\"\nhuman-writing-skills build --style self-media --original original.md --task \"Rewrite without adding facts.\"\nhuman-writing-skills audit --draft chapter.md --context ledger.md --profile physical\nhuman-writing-skills audit --draft chapter.md --profile voice\nhuman-writing-skills audit --draft chapter.md --context ledger.md --profile serial\nhuman-writing-skills audit --draft chapters.md --profile momentum\nhuman-writing-skills audit --draft chapters.md --profile recurrence\nhuman-writing-skills audit --draft chapter.md --profile process\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile ending\nhuman-writing-skills audit --draft paper.md --document-type academic-paper --source study.md --profile sources\nhuman-writing-skills audit --draft revised.md --original original.md --profile fidelity\nhuman-writing-skills audit --draft revised.md --original original.md --profile preservation\nhuman-writing-skills pipeline --draft chapter.md --context ledger.md --auto --with-stats --output-dir audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline outline.md --output-dir novel-audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline outline.md --agent-mode deep --output-dir novel-agent-audit\nhuman-writing-skills verify-chunk-audit --package-dir novel-agent-audit\nhuman-writing-skills chunk-audit --draft report-es.md --style news-report --source sources.md --translationese --agent-mode deep --output-dir report-audit\nhuman-writing-mcp --root C:\\writing-project\nhuman-writing-skills lint --draft chapter.md --style fiction\nhuman-writing-skills stats --draft chapter.md --style fiction\nhuman-writing-skills fix --draft chapter.md --preview\nhuman-writing-skills verify --source original.md --candidate revised.md\nhuman-writing-skills verify-fidelity --source original.md --candidate revised.md\nhuman-writing-skills evaluate --cases tests/fixtures/quality/multilingual-benchmark.json\n```\n\nRead `README.md` or `README.zh-CN.md` for user-facing guidance. Read files under\n`docs/` only for the workflow being used. Do not claim detector evasion or infer\nauthorship from stylistic patterns; frame results as editing evidence.\n\nFile v0.15.3:plugins/deepseek-harness/README.md\n\n# Advanced Human Writing for DeepSeek Harness\n\nThis DeepSeek Harness bundle mounts the `human-writing-mcp` tools from the\nparent repository. It coordinates bounded long-form reviews locally: create a\nplan, let separate agents claim focused work, require complete coverage\nreceipts, and unlock the final reconciliation only after verification passes.\n\n## Prerequisites\n\nInstall the Python package in the environment selected by `PYTHON` or `python`:\n\n```powershell\npip install human-writing-skills\n```\n\nFor development from this repository:\n\n```powershell\npip install -e .\n```\n\n## Install\n\nAfter publishing this folder to npm, add it to a DeepSeek Harness profile:\n\n```powershell\ndsh plugin --profile web add dsh-advanced-human-writing\n```\n\nRestart the affected DSH profile. The bundle uses the workspace as its allowed\nfile root. Set `PYTHON` before launching DSH when the desired interpreter is\nnot named `python`.\n\nThe mounted tools are `plan_long_form_audit`, `list_audit_tasks`,\n`claim_audit_task`, `submit_audit_report`, `verify_audit_coverage`,\n`get_reconciliation_task`, and `read_project_context`.\n\nThis package does not send drafts to a third party and does not call a model by\nitself. The active DSH agent performs each assigned review and must submit its\nown report.\n\nFile v0.15.3:README.md\n\n# Advanced Human Writing & AI Humanizer\n\n> Reusable multilingual writing `SKILLS` for natural prose, genre-aware style, and long-form continuity.\n\n**Advanced AI humanizer and de-AI writing toolkit** for natural rewriting,\nAI text cleanup, fiction editing, novel continuation, chunked long-form audit,\nwriting style unification, and character consistency review.\n\n[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)\n[![Python](https://img.shields.io/badge/Python-3.9%2B-blue.svg)](pyproject.toml)\n[![Zero Dependencies](https://img.shields.io/badge/dependencies-zero-brightgreen.svg)](pyproject.toml)\n\n[中文说明](README.zh-CN.md) | English | [Español](README.es.md) | [Português (Brasil)](README.pt-BR.md) | [Français](README.fr.md)\n\nAdvanced Human Writing & AI Humanizer is an open-source, modular skill pack and lightweight prompt compiler for natural multilingual AI-assisted writing. The package, repository, and ClawHub slug remain `human-writing-skills` for compatibility.\n\nThis is not an empty prompt collection. The repository contains two deliberately\ndifferent capability layers:\n\n- **Executable Python tools:** `humanwriting/` provides the installable\n  `human-writing-skills` CLI for deterministic lint findings with evidence spans,\n  text statistics, conservative fix previews, protected-content verification,\n  prompt compilation, and staged audit-file generation.\n- **Model-executed editorial modules:** `skills/*.md` contains genre and review\n  instructions selected on demand by that compiler. These modules guide the writing\n  model; they do not falsely present subjective literary judgment as a deterministic\n  algorithm.\n\nThe test suite exercises both the executable layer and module-selection gates.\n\n## Agent Orchestration, MCP, And DeepSeek Harness Plugin\n\nFor book-length novels, report series, or research coverage that cannot be\ntrusted to one chat window, install the dependency-free `human-writing-mcp`\nserver. It makes the long-form task graph executable across Codex, Claude Code,\nOpenCode, DeepSeek Harness, Manus, and Hermes: agents claim bounded tasks,\nsubmit coverage receipts, and cannot unlock reconciliation until every required\nreview is complete. The service stays inside a chosen project root and does not\ncall a model or upload drafts.\n\n```powershell\nhuman-writing-mcp --root C:\\writing-project\n```\n\nThe repository also ships a native DeepSeek Harness npm/Cordis bundle under\n[`plugins/deepseek-harness`](plugins/deepseek-harness). It mounts the official\nDSH MCP client and starts the same verified local coordination service. See the\n[agent orchestration guide](docs/agent-orchestration.md) and\n[DeepSeek Harness plugin guide](plugins/deepseek-harness/README.md).\n\nThe MCP server also supports day-to-day single-document work: `lint_text`,\n`get_style_statistics`, `verify_protected_content`, `verify_fidelity`, `compile_humanize_prompt`,\n`compile_audit_prompt`, and `compile_ledger_extraction`. It advertises a small\nnative prompt menu (`humanize-quick`, `dialogue-audit`, `continuity-audit`,\n`serious-rewrite`, and more) for hosts that implement MCP Prompts. These tools\nstay local and return evidence or compiled instructions; they never send a draft\nto a model themselves.\n\n## Ledger Auto-Extraction And Project Defaults\n\nLong-form continuity should not require hand-maintaining every fact from scratch.\n`extract-ledger` compiles an evidence-first prompt for turning existing chapters\ninto a **candidate** continuity ledger. Its output distinguishes observed facts,\ninferences, conflicts, and unknowns, and requires a quote or location for every\nproposed state, object, obligation, injury, resource change, or spatial relation.\nReview it before making it canonical.\n\n```powershell\nhuman-writing-skills extract-ledger --draft chapters-01-10.md --context novel-ledger.md --output ledger-extraction-prompt.md\n```\n\nUse `.humanwriting.json` in a project root to persist only lightweight defaults:\nstyle, document type, a relative ledger path, and lint allow-list entries. Explicit\nCLI flags win, and the file cannot silently activate deep profiles or expensive\nreference/source passes. See the [ledger extraction guide](docs/ledger-extraction.md).\nPass `--no-project-config` when a one-off command must ignore the nearest project file.\n\n```json\n{\n  \"style\": \"fiction\",\n  \"document_type\": \"fiction\",\n  \"context\": \"novel-ledger.md\",\n  \"allow\": [\"END001\"]\n}\n```\n\n## Editor, CI, And Python Distribution\n\nText-facing commands accept `--draft -` for standard input. `lint --format github`\nemits GitHub Actions annotations, while `list --kind rule` prints the current rule\ncatalog with severity, category, and repair direction.\n\n```powershell\ngit diff -- docs\\ | human-writing-skills lint --draft - --style general --format github --source-name docs-change.md\nhuman-writing-skills list --kind rule --format json\n```\n\nEach GitHub Release builds a wheel and source distribution, then attaches them as\nRelease assets. PyPI publication uses GitHub Trusted Publishing only after the\nrepository variable `PYPI_PUBLISH_ENABLED=true` and the PyPI trusted publisher\nhave been configured; no PyPI token is stored in this repository.\n\nIt helps a writing agent move away from generic, template-shaped output and toward prose that has intention, texture, continuity, and genre discipline. The project is especially useful for long-form generation, where characters, settings, arguments, facts, and unresolved threads often drift after several passages.\n\nThe goal is not deception. The goal is better writing: clearer instructions, stronger revision habits, and reusable style constraints that make AI-assisted drafts feel edited by a human.\n\n## Long-Form Audit And Style Unification\n\nOptional `chunk-audit --book-level` adds one whole-book architecture pass after\nchunk reports for fiction/webnovels with at least three chapter headings. It\nchecks consequential choices, conflict development, time movement, and\nover-explained ambiguity. It is not loaded into every chunk or serious report.\n\nThe executable `chunk-audit` workflow splits a year-long novel, article series, or\nlarge report at natural boundaries. Each body span is audited once, with a small\nread-only lead-in and the same user-confirmed style baseline plus outline or project ledger.\nIt writes independent chunk prompts, deterministic cross-chunk style diagnostics, and\na reconciliation prompt for model- or prompt-version drift in narration, character\ndialogue, terminology, and section function.\n\nFor fiction, `--outline` or `--context` makes supported goals, knowledge, relationships,\nlimits, abilities, and speaker voice canonical. Without it, character inferences remain\nprovisional. News, academic, official, and report workflows instead align terminology,\nfacts, attribution, claim scope, and section purpose without loading fiction rules.\n\n```powershell\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline novel-outline.md --output-dir novel-audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline novel-outline.md --book-level --output-dir book-audit\n```\n\nDiscovery terms: **long-form audit, chunked manuscript audit, writing style\nunification, style consistency review, character consistency audit, cross-chapter\ncontinuity, novel audit, and report review**. See the\n[long-form consistency guide](docs/long-form-consistency.md).\n\n## Earned Scene And Document Endings\n\nAI-assisted drafts often reach a real stopping point and then append a scenic dissolve,\nshared silence, life lesson, future-facing reflection, or summary of what the scene\nalready showed. The `earned-ending-audit` finds this **reflective bookend / false\nclosure** pattern in Chinese and English by locating the last meaningful change and\napplying a deletion test. It does not ban sunsets, silence, reflection, or lyrical prose.\n\n- Fiction and webnovels stop on an earned consequence, decision, discovery, changed\n  object, live pressure, or image whose meaning changed inside the scene.\n- Hard news ends on the last useful verified fact, response, constraint, or next step;\n  feature kickers must add meaning instead of manufacturing uplift.\n- Academic, technical, and official writing ends with supported findings, limits,\n  implications, decisions, owners, or deadlines rather than a ceremonial conclusion.\n\n```powershell\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile ending\nhuman-writing-skills lint --draft chapter.md --style fiction\n```\n\nThe full module loads only through the explicit `ending` profile, while `END001` provides a lightweight\ndeterministic preflight for narrative endings. Discovery terms: **AI story ending,\nreflective ending, scene ending audit, chapter ending audit, formulaic conclusion,\nfalse closure, AI reflective bookend, and can't-help-but-reflect ending**.\n\n## Repetition, Exposition, And Scene Economy\n\nSome drafts avoid obvious stock phrases yet still feel generated because they repeat a\nline, image, action sequence, or narrator explanation with no changed consequence.\nThe narrative-only `repetition-exposition-audit` separates intentional refrains and\nneeded recap from exact echoes, repeated choreography, inventory-like viewpoint scans,\nand explanation that merely tells the reader what dialogue or action already showed.\n\n```powershell\n# Isolated, token-bounded narrative pass\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile repetition\n\n# Long drafts: add it only when matching narrative cues appear\nhuman-writing-skills pipeline --draft chapter.md --lint-style fiction --auto --output-dir chapter-audit\n```\n\n`REP001` flags a non-trivial verbatim sentence echo; `NAT005` and `NAT006` are\nlow-severity density prompts for repeated interpretive narration and action frames.\nThey are editing evidence, not authorship claims, and can be allowlisted. Quick\nhumanization and serious-document workflows do not load this module.\n\n## Why This Exists\n\nAI writing often fails in predictable ways:\n\n| Problem | What this project adds |\n| --- | --- |\n| Generic \"AI voice\" | Concrete revision checks for rhythm, specificity, and empty phrasing |\n| Repeated not-X/is-Y, is-X/not-Y, or chained Chinese 比 frames | Family- and density-based checks that preserve necessary correction and real comparison |\n| Rewriting silently changes facts, uncertainty, or causal meaning | An opt-in original-text fidelity pass with a claim ledger and invention checks |\n| Humanizing washes out hesitation, motifs, subtext, or speaker identity | A source-backed preservation ledger that separates useful ambiguity from real defects |\n| Fluent-looking sentences drop a word, object, or connector clause | A separate final pass over predicate slots, parallel structure, and references |\n| Surface AI patterns recur across a passage | Genre-aware checks for vague attribution, inflated significance, false ranges, synonym cycling, formatting habits, and comparison ladders |\n| Fiction is chopped up by time/place mini-headings | Narrative-only checks that preserve titles and chapters but require scene changes to move through prose |\n| A finished scene grows a scenic, reflective, or moralizing tail | Last-meaningful-change and deletion tests for false closure, with genre-specific ending contracts |\n| A scene replays the same action, reaction, or explanation until it becomes mechanical | Narrative-only evidence for exact echoes, repeated action frames, viewpoint inventory, and redundant gloss |\n| One style fits every genre | Separate Markdown `SKILLS` for different writing forms |\n| Long text loses continuity | A compact ledger for facts, plot, promises, and voice anchors |\n| Prose and character dialogue drift across months or model versions | A fixed baseline, canonical outline, unique audit chunks, and cross-chunk reconciliation |\n| Dialogue sounds interchangeable or out of character | Generation and review against baseline voice, scene goal, knowledge, audience, and pressure |\n| Dialogue ends in a stock gesture or scenic gloss instead of a real exchange | A dialogue-performance pass checks listener uptake, purposeful physical beats, and the changed option or carried debt |\n| Dialect, honorifics, particles, or foreign language jump between characters | An evidence-backed language-identity card with motivated switch gates |\n| A consequential line or action receives no uptake before the prose cuts away | Response-obligation checks and deferred interaction debt |\n| Power, skill, authority, equipment, injury, or resources drift | Permanent/temporary state separation and earned transition gates |\n| Prompts become messy | A CLI that compiles style, context, and task into one clean instruction pack |\n| Advice stays abstract | Rules are written as observable editing actions |\n\n## Built-In Style Skills\n\n| Skill | Use it for | Main focus |\n| --- | --- | --- |\n| `fiction` | literary or commercial fiction | point of view, scene pressure, character behavior |\n| `argumentative` | essays and opinion pieces | thesis, evidence, counterargument, logical flow |\n| `news-report` | news-style reports | factual order, attribution, neutral wording |\n| `self-media` | social posts and creator essays | useful voice without empty hype |\n| `academic-paper` | research writing | cautious claims, structure, terminology |\n| `formal-document` | official and administrative documents | authority, scope, responsibility, action, deadline, restrained register |\n| `webnovel` | serialized genre fiction | hooks, payoffs, power rules, continuity |\n\n## Deep Human-Trace Modules\n\nThese modules target deeper AI-writing artifacts, not only surface phrases.\n\n| Module | What it repairs |\n| --- | --- |\n| `controlled-drift` | overly smooth logic, no associative movement, no unfinished thought |\n| `narrative-bridges` | weak scene turns, generic transitions, paragraphs that do not cause each other |\n| `relationship-state` | relationships that reset, dialogue without leverage, forgotten secrets or boundaries |\n| `relationship-stance-audit` | audience-specific stance checks for rivalries, affairs, factions, hierarchy, sects, and family politics |\n| `logic-causality-audit` | cause, timeline, knowledge, motive, rule, resource, and consequence failures |\n| `character-consistency-audit` | character goal, voice, competence, boundary, knowledge, and change-gate drift |\n| `dialogue-voice-audit` | character-fit dialogue plus verbal, physical, silent, interrupted, or deferred uptake for consequential turns |\n| `dialogue-performance-audit` | uses the voice audit's response map to test whether a physical beat changes the exchange rather than decorating it |\n| `speech-register-continuity` | evidence-backed language, dialect exposure, honorifics, particles, address, and switch gates |\n| `capability-state-audit` | power, skill, authority, equipment, injury, resources, cooldowns, counters, and transitions |\n| `serial-reentry` | recap dumps and chapter resets when prior chapters or a ledger are supplied |\n| `long-form-style-consistency` | chunked long-form style, character-setting, and speaker-voice reconciliation |\n| `chapter-momentum-audit` | atmosphere-only chapters, missing payoffs, discarded residue, and unsupported hooks |\n| `world-ontology-audit` | incompatible era, technology, institution, social practice, or speculative rule |\n| `process-earnedness-audit` | promised processes skipped before an unsupported result |\n| `attention-budget-audit` | low-value expansion and semantic echoes displacing consequential material |\n| `chapter-pattern-audit` | repeated chapter architecture across three or more chapters |\n| `story-architecture-audit` | opt-in whole-book decisions, escalation, temporal shape, and ambiguity with chapter evidence |\n| `narrative-distance-control` | unmotivated zoom, missing orientation, and viewpoint-distance drift |\n| `imagery-load-audit` | stacked comparisons, competing sensory channels, and show-then-gloss repetition |\n| `paragraph-rhythm-audit` | mechanical one-line paragraph runs and overloaded long blocks |\n| `detail-disclosure-audit` | biography and appearance inventories delivered before the scene uses them |\n| `scene-entry-audit` | exact-time/location/weather/outfit opening bundles before pressure-bearing action |\n| `natural-measurement` | false precision: tiny exact measures and counted micro-actions in narrative prose |\n| `cliche-phrase-audit` | stock phrases, generic body cues, empty emotion labels, and dead transitions |\n| `formulaic-structure-audit` | triplets, bidirectional contrast frames, chained comparisons, and overly neat closure |\n| `prose-progress-audit` | static paragraphs and pressure-bearing interactions abandoned before uptake or explicit deferral |\n| `narrative-naturalness-audit` | in deep or explicit AI-trace review, catches recurring six-beat scene recipes and copied entry/closure cadence |\n| `repetition-exposition-audit` | narrative-only exact echoes, repeated action choreography, inventory drift, and redundant explanatory narration |\n| `earned-ending-audit` | reflective bookends, scenic dissolves, false closure, stock kickers, and conclusions added after the last meaningful change |\n| `imperfect-prose` | prose that is too clean, too symmetrical, or too polished |\n| `vocal-rhythm` | flat cadence and missing read-aloud breath points |\n| `embodied-emotion` | emotion labels without body, action, contradiction, or perception |\n| `cultural-anchors` | vacuum prose with no era, place, community, or material detail |\n| `spatial-blocking` | character teleportation and confused front/back/left/right blocking |\n| `occupancy-capacity` | over-occupied or mode-ambiguous seats, benches, beds, stools, aisles, and surfaces |\n| `appearance-prop-continuity` | clothing, shoes, props, injuries, and daily-detail drift |\n| `physical-continuity-audit` | optional light manual checklist; do not combine with the forensic physical profile |\n| `proofreading-audit` | final omissions, predicate slots, stranded connectors, references, punctuation, naming, and layout |\n| `style-matrix` | the mistake of applying one generic \"human voice\" to every genre |\n| `editor-loop` | one-shot drafting without a critical human-editor pass |\n| `ai-trace-rubric` | vague feedback like \"sounds AI\" without diagnosis |\n| `reference-style-alignment` | explicit reference material into transferable voice features without copying content |\n| `rewrite-fidelity` | meaning drift, invented specificity, reversed polarity, and altered uncertainty when an original is supplied |\n| `voice-ambiguity-preservation` | over-clean rewrites that erase useful ambiguity, repetition, motifs, hesitation, subtext, or speaker markers |\n| `humanize-examples` | an explicit-only before/after repair library; never loaded as a source or default style sample |\n| `surface-pattern-audit` | recurrent formatting, false ranges, synonym cycling, and narrative mini-headings without global bans |\n| `protected-content` | accidental changes to numbers, citations, equations, URLs, code, quotes, and required terms |\n| `source-grounding` | claim-to-source checks for serious documents with explicit factual sources |\n\n## Quick Start\n\n```powershell\ngit clone https://github.com/whh110112/human-writing-skills.git\ncd human-writing-skills\npython -m pip install .\n\nhuman-writing-skills list --kind style\nhuman-writing-skills list --kind module\nhuman-writing-skills build --style fiction --context examples/story-ledger.md --task \"Write the next scene.\"\nhuman-writing-skills humanize --draft chapter.md --style fiction --mode quick\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline novel-outline.md --output-dir novel-audit\nhuman-writing-skills lint --draft chapter.md --style fiction\nhuman-writing-skills verify --source original.md --candidate revised.md\n```\n\nYou can also run directly from the source checkout with `python -m humanwriting.cli ...`. The `build` and `humanize` commands print instruction packs that can be pasted into Codex, ChatGPT, Claude, local LLM tools, or another writing agent.\n\n## Quick Humanize\n\n`humanize` is the low-friction rewrite route. It treats `--draft` as the original,\nkeeps the same language and genre by default, and preserves meaning before changing\nsurface style.\n\n```powershell\n# Minimum stack: surface patterns + fidelity + voice/ambiguity preservation\nhuman-writing-skills humanize --draft chapter.md --style fiction\n\n# Add structural editor passes only when the draft needs them\nhuman-writing-skills humanize --draft article.md --style self-media --mode deep\n\n# Examples remain opt-in and are never treated as factual or stylistic source material\nhuman-writing-skills humanize --draft chapter.md --style fiction --with-examples\n```\n\n`quick` does not load cliche, formulaic-structure, paragraph-progress, or editor-loop\nmodules. `deep` adds those high-cost passes. `humanize-examples` loads only with\n`--with-examples`; `voice-ambiguity-preservation` loads only for supplied-text\nhumanization or an explicit preservation audit.\n\n## Multilingual Scope\n\nThe skill instructions have no Chinese-only gate: they can guide fiction and serious\nprose in English, Japanese, French, Spanish, Portuguese, Arabic, Latin, and other\nlanguages supported by the selected model. Deterministic lexical rules are naturally\nlanguage-specific, while structural continuity and review remain language-agnostic.\nThe narrative heading scanner recognizes time cards across the languages above, and\n`stats` profiles Han, kana, Arabic, and several Latin-script language families. Use\ngenre context and human review for mixed-language or low-resource text.\n\nUntyped serious-document detection uses multiple language-specific cues for\nFrench, Spanish, Portuguese, Japanese, and Arabic as well as Chinese and English.\nSet `--document-type` explicitly when cues are sparse. Non-Chinese, non-English\nstyle statistics remain descriptive, without purportedly validated thresholds.\n\n## Quality Evaluation\n\nDeveloper-written multilingual seed cases test routing false positives/negatives\nand rewrite-fidelity triage. They are regression examples, not native-speaker\nvalidation. `evaluate` reports separately by language, genre, and reviewer status:\n\n```powershell\nhuman-writing-skills evaluate --cases tests/fixtures/quality/multilingual-benchmark.json\n```\n\nSee the [annotation protocol](docs/evaluation.md) before making accuracy claims.\n\n## Example Output Shape\n\n```text\n# Core Directive\n# Continuity Protocol\n# Selected Skill: fiction\n# Project Context\n# Task\n# Output Contract\n```\n\nThis format keeps the model focused on the current task while still carrying the previous facts, style decisions, and unresolved threads.\n\n## Explicit Reference Style\n\nReference matching is opt-in. It activates only with `--reference`,\n`--reference-style`, or explicit task wording such as \"match this voice.\" A\ncontinuity ledger by itself never activates it.\n\n```powershell\nhuman-writing-skills build `\n  --style fiction `\n  --context examples/story-ledger.md `\n  --reference examples/reference-style-source.zh-CN.md `\n  --task \"Continue the scene while matching the reference's restrained rhythm.\"\n\nhuman-writing-skills audit `\n  --draft my-chapter.md `\n  --reference examples/reference-style-source.zh-CN.md `\n  --profile style-match\n```\n\nThe compiler extracts point of view, rhythm, register, imagery, description,\ndialogue cadence, emotion handling, and transitions. Plot facts still come from\n`--context`; names, events, and distinctive phrases must not be copied from the\nreference. See [docs/reference-style.md](docs/reference-style.md).\n\n## Original-Text Fidelity\n\n`verify` checks only exact protected tokens. For consequential revisions, also\nrun `verify-fidelity --source original.md --candidate revised.md`. It locates\nchanged claim spans and potential actor, negation, uncertainty, scope, or\nattribution changes. Changed prose is **needs review**, never an automatic\nsemantic pass. Exit codes: 0 identical, 1 literal mismatch, 2 changed prose\nrequiring claim review. See [protected-content guidance](docs/protected-content.md).\n\nUse `--original` only when revising an existing text and meaning must remain stable.\nIt activates a dedicated fidelity module for rewrite and review; ordinary drafting\ndoes not pay this token cost.\n\n```powershell\nhuman-writing-skills build `\n  --style self-media `\n  --original original.md `\n  --task \"Rewrite for clarity without adding facts or strengthening claims.\"\n\nhuman-writing-skills audit `\n  --draft revised.md `\n  --original original.md `\n  --profile fidelity\n```\n\n`--original` is semantic authority, `--reference` is style evidence, and `--source`\nis factual evidence for serious documents. They are deliberately isolated so a\nstyle sample cannot rewrite facts and an original cannot silently become a style\ntarget. See [docs/editing-tools.md](docs/editing-tools.md).\n\n## Serious-Document Sources\n\n`--source` is separate from `--reference`. It activates `source-grounding` only for\nacademic, news, legal, or technical work and builds a claim-to-source evidence map.\nFiction, webnovels, self-media, and casual answers do not auto-load it.\n\n```powershell\nhuman-writing-skills audit `\n  --draft paper.md `\n  --document-type academic-paper `\n  --source study-a.md `\n  --source study-b.md `\n  --profile sources\n```\n\nThe audit separates source existence from claim support. Without external registry\naccess, it marks citation metadata as unverified instead of inventing a verdict.\n\n## Long-Form Continuity\n\nFor longer works, this project recommends a small ledger instead of relying only on a large context window. Use context in this order: canonical ledger, latest confirmed state, recent chapters, relevant retrieved older spans, then explicitly uncertain inference. Retrieved text is recall evidence and cannot overwrite a later canonical state.\n\nThe ledger tracks:\n\n- fixed facts: names, dates, locations, relationships, rules, timeline\n- active threads: unresolved conflicts, clues, promises, open arguments\n- relationship state: who knows, wants, hides, owes, refuses, or holds leverage\n- relationship stance: public/private posture, current audience, mention policy, forbidden leaks, and exception motives\n- voice anchors: point of view, diction, directness, disclosure habits, domain limits, audience shifts, taboo phrases\n- language identity: shared scene language, demonstrated dialect/second-language exposure, address forms, particles, and switch gates\n- capability state: permanent power/skill/authority plus temporary injury, equipment, resources, cooldowns, counters, costs, and transition gates\n- dialogue contract: who speaks to whom, why now, desired listener action, protected information, and intended state change\n- interaction debt: which consequential line or action still awaits uptake, refusal, interruption, consequence, or delayed payoff\n- current state: where the previous passage ended and what must connect next\n- beat bridge: previous residue, entry pressure, micro-turn, and exit hook\n- change log: what became newly true in the latest output\n\nSee [examples/story-ledger.md](examples/story-ledger.md) for a fiction example.\n\n`speech-register-continuity` auto-loads only for fiction/webnovel dialogue when the task or ledger contains explicit language, regional, dialect, honorific, or register evidence. It can also be selected with `audit --profile register`; region or nationality never licenses an invented accent.\n\n`capability-state-audit` loads during generation only when the current task names a capability constraint. Automatic pipeline review additionally requires context, so ordinary dialogue scenes do not pay its Token cost. Select it explicitly with `audit --profile capability --context ledger.md` when needed.\n\n## Chatbox\n\nYes, this project works in Chatbox because it outputs plain text prompt packs. For long writing sessions, use the continuity ledger as the source of truth and paste the compiled prompt pack into Chatbox's system prompt or first message.\n\n- English guide: [docs/chatbox.md](docs/chatbox.md)\n- Chinese guide: [docs/chatbox.zh-CN.md](docs/chatbox.zh-CN.md)\n- Ledger template: [examples/chatbox-ledger-template.md](examples/chatbox-ledger-template.md)\n\n## Physical Continuity\n\nFor scenes where space matters, such as cars, elevators, hospital rooms, dining tables, and bedrooms, use `--strict-continuity`. It adds occupancy, spatial blocking, and appearance/prop generation guards. Use `audit --profile physical` for one evidence-first forensic pass on an existing draft; it owns capacity, blocking, appearance, props, barriers, reach, and body-state contradictions in one ledger.\n\n```powershell\npython -m humanwriting.cli build `\n  --style fiction `\n  --strict-continuity `\n  --review `\n  --context examples/vehicle-scene-ledger.md `\n  --task \"Continue the car argument. Every seat change must have an on-page transition. Keep clothing and props consistent.\"\n```\n\n- Guide: [docs/physical-continuity.md](docs/physical-continuity.md)\n- Vehicle ledger example: [examples/vehicle-scene-ledger.md](examples/vehicle-scene-ledger.md)\n- Capacity ledger template: [examples/capacity-ledger-template.md](examples/capacity-ledger-template.md)\n- Capacity conflict example: [examples/capacity-conflict-draft.zh-CN.md](examples/capacity-conflict-draft.zh-CN.md)\n- Draft audit example: [examples/problem-car-scene-draft.md](examples/problem-car-scene-draft.md)\n\n## Relationship Stance Continuity\n\nFor scenes with rival factions, secret relationships, hierarchy, family politics,\noffice politics, or sect leaders, use `--deep-review` or add `relationship-stance-audit`.\nIt extracts each dialogue line as `speaker -> listener/audience -> referenced party`\nand checks whether praise, criticism, comparison, naming, secrecy, and rank fit\nthe established relationship graph.\n\n- Guide: [docs/relationship-stance-continuity.md](docs/relationship-stance-continuity.md)\n- Ledger template: [examples/relationship-stance-ledger.zh-CN.md](examples/relationship-stance-ledger.zh-CN.md)\n\n## Character- and Situation-Fit Dialogue\n\n**Bidirectional interaction review addresses missing reactions and unfinished actions.**\nIt covers speech, wordless exchanges, and the viewpoint character's own responses.\nTrack initiator, affected recipient, separate speech/action obligations, evidence of\nreception, and completed or pending state. A spoken answer may leave a simultaneous\naction unfinished; one person's movement does not establish that another followed.\nRefusal, unawareness, stillness, or delayed uptake can be valid. Clearly implied\ncompletion needs no added gesture or emotional commentary.\n\nFiction and webnovel base Skills retain a brief check; detailed review uses the\nexisting on-demand `voice` profile. Automatic pipelines and deep chunk reviews also\nroute directed interpersonal actions. Use `--profile voice` when automatic cues miss\na language or phrasing. Reconciliation checks responses across chunk boundaries.\nThese are model-executed contextual checks; routing tests and deterministic lint do\nnot establish that a model will catch every omitted reaction.\n\n`dialogue-voice-audit` and `dialogue-performance-audit` separate stable speaker baseline,\nsituation-driven modulation, and the action each turn is trying to perform. Occupation, class, region, and trait\nlabels supply possible knowledge, incentives, duties, and register pressure; they do\nnot substitute for personality. An explicit speech or interpersonal-action task activates\nthe module on demand. Review an existing scene with an independent `voice` pass:\n\n```powershell\nhuman-writing-skills audit `\n  --draft my-dialogue-scene.md `\n  --context my-novel-ledger.md `\n  --profile voice\n```\n\nThe audit separates contradiction from motivated contrast and checks scene purpose,\nknowledge boundaries, practical constraints, response linkage, audience, and power.\nA consequential line or action does not require a mechanical spoken reply, but it\nmust receive verbal, physical, silently legible, interrupted, or deliberately deferred\nuptake before the prose shifts away. A physical beat is retained only when it changes\naccess, attention, leverage, permission, distance, or the next available action; the\nmodule does not prescribe touch, gestures, weather, clothing, or scenery after every line.\n\nIf the draft already exists, use `audit`:\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft examples/problem-car-scene-draft.md `\n  --context examples/vehicle-scene-ledger.md\n```\n\n## Project Layout\n\n```text\nhumanwriting/        Python package and CLI\nskills/              reusable writing SKILLS in Markdown\nexamples/            sample continuity ledgers and article briefs\ntests/               standard-library unit tests\n```\n\n## CLI Usage\n\n### Optional Narrative Modules\n\nThe narrative controls use progressive disclosure. Generation adds\nthe dialogue modules only when a fiction or webnovel task explicitly asks for a\nspeech-centered or character-interaction scene such as dialogue, negotiation, reunion,\ntesting, reconciliation, confrontation, a meeting, interrogation, or argument.\nNarration-only and serious-document tasks do not trigger them. The `voice`,\n`serial`, `world`, `process`, `momentum`, `salience`, `recurrence`, `repetition`, `texture`, and\n`sources` and `preservation` audit profiles remain outside broad `full` review:\n\n```powershell\nhuman-writing-skills build --style fiction --task \"Write a negotiation in which both speakers want different outcomes.\"\nhuman-writing-skills build --style webnovel --context ledger.md --module serial-reentry --task \"Continue chapter 18.\"\nhuman-writing-skills audit --draft chapters.md --profile momentum\nhuman-writing-skills audit --draft chapter.md --profile texture\nhuman-writing-skills audit --draft chapter.md --profile process\nhuman-writing-skills audit --draft chapters.md --profile recurrence\n```\n\nThe dialogue modules model baseline speech, practical incentives, knowledge limits,\nscene goals, response linkage, and purposeful performance beats without treating a job as a personality. `dialogue-voice-audit` owns response obligation and deferred interaction debt; `dialogue-performance-audit` only tests whether a selected physical beat earns its place. Use\n`speech-register-continuity` for evidence-backed language identity, particles,\nhonorifics, and code-switching; use `capability-state-audit` for power and resource\nstate. Use `serial-reentry` only with\nprior chapters or a ledger, `momentum` for a multi-chapter draft, and `texture` for\nnarrative distance, cinematic opening stacks, imagery load, paragraph fragmentation,\nemotional over-explanation, and detail inventory. Use `world` only with explicit\nsetting constraints, `process` for consequential domain work, `salience` for long\ndrafts, `recurrence` for at least three chapters, `repetition` for long enough narrative\ndrafts with matching repetition/exposition cues, and `sources` only with serious\ndocuments and factual source files.\n\nDuring generation, world, process, and attention-budget modules activate only from\nexplicit setting, consequential-process, expansion, long-form, or dilution signals;\nordinary `--deep-review` does not load them.\nThe repetition/exposition module is likewise audit-first: use `--profile repetition`\nor narrative `ai-trace`, and let `pipeline --auto` add its separate pass only when\nthe draft supplies matching cues.\n\n### Audit Profiles\n\n`audit` can load only the checks needed for the current pass:\n\n| Profile | Purpose |\n| --- | --- |\n| `full` | Broad default audit; high-cost and strongly gated profiles remain separate |\n| `logic` | Cause, timeline, knowledge, motive, rules, resources, and consequences |\n| `character` | Character goal, voice, competence, boundaries, and change gates |\n| `voice` | Speaker baseline, scene goal, role/knowledge limits, audience register, change gates, and response obligations |\n| `register` | Language identity, dialect exposure, honorifics, particles, vocabulary, and code-switch gates |\n| `capability` | Power, skill, authority, equipment, injury, resources, counters, and transition gates; requires `--context` |\n| `serial` | Recap dumps, missing carryovers, and chapter resets; requires `--context` |\n| `momentum` | Multi-chapter entry pressure, irreversible turns, payoff, residue, and exit pressure |\n| `world` | Era, technology, institution, social-practice, and world-rule compatibility |\n| `process` | Promise, attempt, resistance, judgment, cost, evidence, and earned result |\n| `salience` | Long-draft attention allocation, dilution, and semantic echoes |\n| `recurrence` | Chapter fingerprints and repeated architecture across three or more chapters |\n| `repetition` | Narrative-only exact echoes, repeated action frames, and explanation that duplicates visible evidence |\n| `texture` | Narrative distance, scene-entry load, imagery, paragraph cadence, and detail disclosure |\n| `physical` | Position, capacity, reach, clothing, props, and injuries |\n| `relationship` | Audience, stance, information permissions, rank, and secret leaks |\n| `ai-trace` | Cliches, formulaic structure, static paragraphs, and other AI traces |\n| `ending` | Last meaningful change, reflective bookends, false closure, and genre-specific ending function |\n| `numbers` | False precision in action and emotion |\n| `proofread` | Omissions, sentence slots, stranded connectors, references, punctuation, naming, and layout |\n| `fidelity` | Meaning, entity, polarity, uncertainty, chronology, attribution, and invention checks; requires `--original` |\n| `preservation` | Useful ambiguity, repetition, motifs, hesitation, subtext, and speaker identity; requires `--original` and explicit selection |\n| `style-match` | Drift from explicitly supplied reference material; unavailable without a reference signal |\n| `sources` | Claim grounding against factual sources; requires a serious document and `--source` |\n\nProfiles can be combined, for example `--profile relationship --profile ai-trace`.\n\nOrdinary generation loads only a lightweight sentence-completeness guard. Full\nomission, missing-object, stranded-connector, and reference checks load only in the\n`proofread` profile or pipeline proofreading stage, preserving the generation token budget.\n\n### Chunked Long-Form Audit\n\nUse `chunk-audit` when a manuscript exceeds one reliable context window or was written\nacross model, prompt, or time changes. It complements `pipeline`: chunking handles\nmanuscript size and cross-block drift, while the pipeline separates different review\nresponsibilities for one draft.\n\n```powershell\nhuman-writing-skills chunk-audit `\n  --draft full-novel.md `\n  --style fiction `\n  --outline novel-outline.md `\n  --reference approved-sample.md `\n  --output-dir novel-consistency-audit\n```\n\nWithout an explicit reference, `--baseline-chunk` selects a candidate manuscript block;\napprove or correct it during baseline extraction. Reference prose supplies style evidence only. Fiction uses the\noutline or ledger for character canon and permits earned development; serious reports\nprotect facts, numbers, terminology, attribution, and conclusion scope. Default body,\ncontext, and baseline budgets keep the workflow usable on smaller-context models.\n\n### Verified Agent Review For Long Documents\n\nAdd `--agent-mode deep` when coverage matters more than token cost. It writes an explicit\ntask graph in `agent-plan.json`, requires a Coverage Receipt from every reviewer, and\nreserves report paths under `reports/`. After the baseline is approved, tasks that depend\nonly on `baseline` may run in parallel in fresh model conversations or API calls.\n\n```powershell\nhuman-writing-skills chunk-audit `\n  --draft full-novel.md `\n  --style fiction `\n  --outline novel-outline.md `\n  --agent-mode deep `\n  --output-dir novel-agent-audit\n\nhuman-writing-skills verify-chunk-audit `\n  --package-dir novel-agent-audit\n```\n\nDo not run `9999-reconcile-prompt.md` until verification reports complete coverage. Deep\nmode adds a paragraph-level prose pass for every block, an interaction pass where dialogue\nor directed actions occur, and an evidence pass only for serious documents with explicit `--source` files.\nStandard mode remains one complete audit per unique block. For an explicitly translated or\nlocalized long document, add `--translationese`; it never activates merely because the text\nuses another language.\n\n- Guide: [docs/long-form-consistency.md](docs/long-form-consistency.md)\n\n### Multi-Stage Pipeline\n\nFor high-precision review, generate independent single-purpose passes instead of asking one model to check everything at once:\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --auto `\n  --output-dir chapter-audit\n```\n\nRun every stage in a fresh model conversation or independent API request. Automatic mode keeps logic, AI-trace, and proofreading stages, then adds focused stages only when their cues and gates match. `serial` and `capability` require context; `fidelity` requires an original; `salience` requires a long narrative of at least 4,000 characters; `recurrence` requires at least three chapters; `repetition` requires a sufficiently long narrative with repeated explanation or choreography cues; and `sources` requires both a serious document and explicit factual sources. The higher-cost `preservation` comparison is explicit-only: use `--stage preservation --original original.md`. Add `--with-stats` only when distributional diagnostics are useful. The manifest explains every selection and skip.\n\n- Guide: [docs/audit-pipeline.md](docs/audit-pipeline.md)\n\n### Deterministic Safeguards\n\nUse `lint` for evidence-located pattern checks, `stats` for optional distributional\ndiagnostics, `fix` for conservative mechanical cleanup, and `verify` to catch\nprotected facts changed during rewriting. Scores and statistics are editing\nheuristics, not authorship proof.\n\nProtected-content instructions auto-load only for academic papers, formal documents,\nnews reports, and strongly identified legal or technical documents. Fiction, webnovels, casual\nQ&A, playful text, and self-media do not auto-load them; use `--protect-content`\nor `--protect-term` to override this gate.\n\nWith `--style fiction` or `--style webnovel`, `lint` also flags unrequested narrative\nmini-headings and multilingual standalone time cards. It preserves work/chapter titles\nand does not apply the rule to news or academic section headings. The repair restores\na prose transition; it does not merely delete the label and join two disconnected blocks.\n\n```powershell\nhuman-writing-skills lint --draft my-chapter.md --style fiction\nhuman-writing-skills stats --draft my-chapter.md --style fiction\nhuman-writing-skills fix --draft my-chapter.md --preview\nhuman-writing-skills verify --source original.md --candidate revised.md --protect-term \"Project Atlas\"\n```\n\n- Pattern lint: [docs/pattern-linter.md](docs/pattern-linter.md)\n- Fidelity, statistics, and conservative fixes: [docs/editing-tools.md](docs/editing-tools.md)\n- Protected content: [docs/protected-content.md](docs/protected-content.md)\n\n### Number Sense\n\nUse this to catch false precision such as unnecessary exact centimeters, seconds, or micro-counts in emotional and bodily action, while preserving necessary numbers in medicine, forensics, engineering, architecture, news, and technical writing.\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft examples/false-precision-draft.zh-CN.md `\n  --profile numbers\n```\n\n- Guide: [docs/number-sense.md](docs/number-sense.md)\n- Example: [examples/false-precision-draft.zh-CN.md](examples/false-precision-draft.zh-CN.md)\n\n### Common Writing Problems\n\nThe project converts recurring long-form writing problems into executable checks: stock phrasing, plastic prose, triplet structures, over-smooth transitions, static paragraphs, hollow emotion, cultural vacuum, and long-form drift.\n\n- Rule map: [docs/forum-complaint-research.md](docs/forum-complaint-research.md)\n\nList styles:\n\n```powershell\npython -m humanwriting.cli list\n```\n\nBuild a prompt pack:\n\n```powershell\npython -m humanwriting.cli build `\n  --style webnovel `\n  --module narrative-bridges `\n  --module relationship-state `\n  --module natural-measurement `\n  --module embodied-emotion `\n  --module vocal-rhythm `\n  --strict-continuity `\n  --review `\n  --context examples/story-ledger.md `\n  --task \"Continue chapter 3. Keep the confrontation unresolved but reveal one new clue.\"\n```\n\nThe compact `--review` flag adds only:\n\n- `editor-loop`: draft, diagnose, locally rewrite, then finalize\n- `ai-trace-rubric`: score cognitive smoothness, generic diction, emotional flatness, rhythm monotony, context drift, weak beat bridges, relationship resets, false precision, cultural vacuum, over-clean prose, and closure addiction\n\nThe `--deep-review` flag adds the compact review plus:\n\n- `relationship-stance-audit`: check speaker, listener, referenced party, secrecy, stance, rank, and audience permissions\n- `cliche-phrase-audit`: check stock phrases, generic body cues, empty emotion labels, and dead transitions\n- `formulaic-structure-audit`: check triplets, bidirectional contrasts, chained comparisons, and paragraphs that close too neatly\n- `surface-pattern-audit`: check recurring significance, attribution, range, lexical, formatting, and mini-heading patterns in genre context\n- `prose-progress-audit`: check whether each paragraph advances facts, relationships, evidence, action, or pressure\n- `narrative-naturalness-audit`: in deep or explicit AI-trace review, check recurring scene-entry/closure recipes across six or more beats; local phrase, structure, and uptake findings stay with their owning modules\n- `natural-measurement`: check false precision in fiction, webnovels, and self-media\n\nThe `--strict-continuity` flag adds:\n\n- `spatial-blocking`: position and movement checks\n- `occupancy-capacity`: physical resource mode, capacity, occupancy, and transformation checks\n- `appearance-prop-continuity`: clothing, shoes, props, and body-state checks\n\nUse `audit --profile physical` for the final physical-state contradiction pass. The optional `physical-continuity-audit` is a light manual checklist and cannot be combined with this forensic profile.\n\nRun tests:\n\n```powershell\npython -m unittest discover -s tests -v\n```\n\n## Writing Philosophy\n\nGood AI-assisted prose should be:\n\n- situated: it knows who is speaking, what changed, and why this passage exists\n- specific: it uses details that belong to this topic, not any topic\n- continuous: it respects previous facts, costs, injuries, claims, and promises\n- shaped: it understands the genre before choosing structure and diction\n- revised: it removes filler, canned transitions, and decorative certainty\n\n## Editorial Guardrails\n\nThis project avoids claiming that any tool can perfectly hide authorship or beat detectors. It focuses on craft: voice, context, genre, revision, and continuity.\n\nWhen studying published work, use short analysis, public-domain sources, licensed material, or your own examples. Do not copy protected passages into skills.\n\n## Contributing\n\nContributions are welcome. Useful additions include:\n\n- new Markdown skills\n- Chinese and multilingual style packs\n- model-specific adapters\n- stronger continuity ledger examples\n- tests for prompt compilation and context preservation\n\nPlease keep each skill practical. A good rule should tell the model what to do, what to avoid, and how to check the result.\n\n## License\n\nMIT. See [LICENSE](LICENSE).\n\nFile v0.15.3:_meta.json\n\n{\n  \"ownerId\": \"kn7cv6zdh5xrvfc30xqepze49582aw5n\",\n  \"slug\": \"human-writing-skills\",\n  \"version\": \"0.15.3\",\n  \"publishedAt\": 1790683340699\n}\n\nFile v0.15.3:CONTRIBUTING.md\n\n# Contributing\n\nThank you for improving Advanced Human Writing & AI Humanizer.\n\n## Rules And Tests\n\n- Add a deterministic rule only when it has a concrete, user-visible editing purpose.\n- Give every rule a stable ID, category, severity, repair direction, allow-list behavior,\n  positive fixture, and a plausible negative fixture.\n- Do not describe a pattern score as proof of AI authorship. Rules identify editing leads.\n- Gate genre-specific rules so fiction, news, academic, legal, and technical writing do\n  not inherit one another's constraints.\n- Preserve token budgets: optional or specialized modules must not silently enter quick\n  generation or unrelated document types.\n\nRun the full suite before opening a pull request:\n\n```powershell\npython -m unittest discover -s tests -v\npython -m build\n```\n\n## Documentation\n\nUpdate the English and Chinese README material for user-facing commands. Keep examples\ngeneric and evidence-led; do not publish private manuscript material in fixtures or issues.\n\n## Pre-commit\n\nTeams can install the repository hook with:\n\n```powershell\npre-commit install --hook-type pre-commit\n```\n\nThe hook checks Markdown-like text and fails only once the transparent pattern score\nreaches its configured threshold. Use allow-lists for intentional motifs or house style.\n\nFile v0.15.3:docs/agent-orchestration.md\n\n# Agent Orchestration And MCP\n\nThis project can run as more than a prompt-only Skill. Its local MCP server\nturns the verified long-form package into a small coordination surface that\nmultiple agents can share without giving every session the whole manuscript.\n\n## What The Server Does\n\n`human-writing-mcp` operates only inside an explicit project root. It does not\ncall a model and it does not upload drafts. It exposes long-form coordination,\nsingle-text analysis, and compiled editorial instructions:\n\n| Tool | Purpose |\n| --- | --- |\n| `plan_long_form_audit` | Build the bounded package and task graph. |\n| `list_audit_tasks` | Show dependencies, claims, receipt validity, and readiness. |\n| `claim_audit_task` | Give one agent one complete, focused prompt. |\n| `submit_audit_report` | Store a report only when its receipt declares complete coverage. |\n| `verify_audit_coverage` | Gate reconciliation on every required report. |\n| `get_reconciliation_task` | Return the cross-document editor prompt after the gate passes. |\n| `read_project_context` | Read a bounded ledger, outline, source note, or approved reference. |\n| `lint_text` | Return local deterministic findings and evidence spans for supplied text. |\n| `get_style_statistics` | Return sentence variation, MATTR, and transition density. |\n| `verify_protected_content` | Compare source and rewritten text for protected values and terms. |\n| `compile_humanize_prompt` / `compile_audit_prompt` | Return a bounded prompt for one supplied passage. |\n| `compile_ledger_extraction` | Return an evidence-backed candidate-ledger extraction prompt. |\n\nThe server rejects file paths outside its root. A receipt must include the task\nID, `Coverage: complete`, checked units, findings, and `Unchecked or blocked\nmaterial: none`. A missing, blocked, or partial receipt keeps reconciliation\nclosed.\n\n## Native Prompt Menu\n\nHosts that implement MCP Prompts can discover a curated menu through `prompts/list`\nand resolve it through `prompts/get`. The server exposes task-level prompts rather\nthan every internal module: `humanize-quick`, `humanize-deep`, `dialogue-audit`,\n`continuity-audit`, `style-match`, `serious-rewrite`, and `extract-ledger`.\nClient UI determines whether these appear as slash commands. Each prompt requires\nthe passage text and may accept compact context; no prompt calls an external model.\n\n## Local Stdio Setup\n\nInstall this project, then configure any MCP-capable host to launch:\n\n```text\ncommand: python\nargs: [\"-m\", \"humanwriting.mcp_server\", \"--root\", \"C:/writing-project\"]\n```\n\nThe project root should contain the manuscript, its outline or continuity\nledger, and generated audit directories. Use project-relative paths in tool\ncalls. The same server works with Codex, Claude Code, OpenCode, DeepSeek\nHarness, and Hermes when they are connected to the same local workspace.\n\n## Remote HTTP Setup\n\nFor a host that only accepts remote MCP endpoints, start the server behind your\nown TLS reverse proxy:\n\n```powershell\nhuman-writing-mcp --root C:\\writing-project --http --host 127.0.0.1 --port 8765 --auth-token <long-random-token>\n```\n\nExpose `POST /mcp` only through HTTPS and preserve the `Authorization: Bearer`\nheader. A non-loopback bind is refused without `--auth-token`. Do not expose a\nmanuscript workspace to the public internet merely to connect an agent.\n\n## Reliable Long-Form Run\n\n1. Call `plan_long_form_audit` with `agent_mode: \"deep\"`.\n2. Have one agent claim and submit `baseline`.\n3. Start fresh agents for the tasks that become ready. Each agent receives only\n   its body, read-only lead-in, compact canon, and specialist modules.\n4. Agents submit reports rather than merely saying that they finished.\n5. Call `verify_audit_coverage`. Repair every missing or invalid receipt.\n6. Call `get_reconciliation_task` and run one final editor only after the gate\n   passes.\n\nThis is deliberately model-neutral. A host may use local subagents, API jobs,\nor separate conversations. The package remains the source of truth, not a\nsingle model's fading chat memory.\n\n## Host Adapters\n\n- **Codex:** Keep a concise `AGENTS.md` beside the writing project, install the\n  Skill for workflow discovery, and add the stdio MCP server for task state.\n- **Claude Code:** Put the matching project contract in `CLAUDE.md`; use its\n  subagents for ready tasks and a hook or final checklist that calls coverage\n  verification.\n- **OpenCode:** It discovers the existing `.agents/skills` format directly;\n  connect the same MCP command for shared queue state.\n- **DeepSeek Harness:** Use the included npm bundle under\n  `plugins/deepseek-harness`. It mounts the official DSH MCP client and starts\n  this server automatically for the active workspace.\n- **Manus:** Import the Skill from GitHub today. For MCP, deploy the HTTP mode\n  on infrastructure you control, behind HTTPS and bearer authentication; its\n  root must contain the intended project files.\n- **Hermes Agent:** Install the Skill from GitHub or a skills tap, then attach\n  the stdio or protected HTTP MCP server. Its subagents can claim separate\n  ready tasks.\n\n`integrations/project-context/AGENTS.md.example` is intentionally short. It\npoints agents to the ledger and MCP workflow rather than duplicating all audit\nrules into every conversation.\n\nFile v0.15.3:docs/agent-orchestration.zh-CN.md\n\n# Agent 编排与 MCP\n\n本项目不只是提示词式 Skill。`human-writing-mcp` 可以把已验证的长篇审查包变成一个小型\n协作界面，让多个 Agent 共用同一份计划、账本与回执，而不必把整部作品塞进每个会话。\n\n## 服务做什么\n\n服务只访问明确指定的项目根目录；不会自行调用模型，也不会上传草稿。它提供长篇协作、\n单篇分析与编辑指令编译工具：\n\n| 工具 | 作用 |\n| --- | --- |\n| `plan_long_form_audit` | 生成分块审查包与任务图。 |\n| `list_audit_tasks` | 查看依赖、领取状态、回执有效性与可执行状态。 |\n| `claim_audit_task` | 把一个完整且有边界的任务交给一名 Agent。 |\n| `submit_audit_report` | 只有回执声明完整覆盖时才保存报告。 |\n| `verify_audit_coverage` | 对所有必需报告做统稿前门禁。 |\n| `get_reconciliation_task` | 门禁通过后才返回跨章统稿任务。 |\n| `read_project_context` | 读取有长度上限的账本、大纲、资料说明或文风参考。 |\n| `lint_text` | 对传入文本返回本地确定性发现项与证据位置。 |\n| `get_style_statistics` | 返回句长变化、MATTR 与转折词密度。 |\n| `verify_protected_content` | 比对原文与改写稿的受保护值和术语。 |\n| `compile_humanize_prompt` / `compile_audit_prompt` | 为单篇文本编译有边界的 Prompt。 |\n| `compile_ledger_extraction` | 编译带证据的待确认账本提取 Prompt。 |\n\n所有越出根目录的路径都会被拒绝。回执必须带任务 ID、`Coverage: complete`、检查单元、发现\n项以及 `Unchecked or blocked material: none`。漏审、阻塞或只审了一部分，均不能进入统稿。\n\n## 原生 Prompt 菜单\n\n支持 MCP Prompts 的客户端可通过 `prompts/list` 发现菜单，并通过 `prompts/get` 获取 Prompt。\n服务提供的是任务级快捷入口，而不是把全部内部模块暴露成冗长列表：`humanize-quick`、\n`humanize-deep`、`dialogue-audit`、`continuity-audit`、`style-match`、`serious-rewrite` 与\n`extract-ledger`。客户端是否将其显示为斜杠命令取决于客户端 UI；所有 Prompt 都要求传入正文，\n不会自行调用外部模型。\n\n## 本地 stdio 接入\n\n先安装本项目，再在支持 MCP 的 Agent 中配置：\n\n```text\ncommand: python\nargs: [\"-m\", \"humanwriting.mcp_server\", \"--root\", \"C:/writing-project\"]\n```\n\n根目录中应包含稿件、人物/世界账本或报告计划，以及生成的审查目录。所有工具调用都使用项目\n相对路径。同一服务可供 Codex、Claude Code、OpenCode、DeepSeek Harness 与 Hermes 共用。\n\n## 远程 HTTP 接入\n\n只接受远程 MCP 的平台，可在自有 HTTPS 反向代理后启动：\n\n```powershell\nhuman-writing-mcp --root C:\\writing-project --http --host 127.0.0.1 --port 8765 --auth-token <long-random-token>\n```\n\n仅通过 HTTPS 暴露 `POST /mcp`，并保留 `Authorization: Bearer` 请求头。非本机地址未提供\n`--auth-token` 时会被拒绝。不要为了连接远程 Agent 而把小说或资料目录裸露在公网。\n\n## 长篇可靠运行流程\n\n1. 用 `agent_mode: \"deep\"` 调用 `plan_long_form_audit`。\n2. 先由一名 Agent 领取并提交 `baseline`。\n3. 对已就绪任务启动独立会话或子 Agent；每个任务只带正文块、只读前文、压缩账本和专项规则。\n4. Agent 必须提交报告，不能只口头声称“已经审完”。\n5. 调用 `verify_audit_coverage`，修复全部缺失或无效回执。\n6. 只有验证通过后，才领取 `get_reconciliation_task` 做跨章统稿。\n\n这套流程不绑定模型：可以是本地子 Agent、API 任务或多段独立会话。真正的权威是项目审查包，\n不是某一个对话窗口逐渐衰减的记忆。\n\n## 平台适配\n\n- **Codex：** 在写作项目旁保留简短 `AGENTS.md`，Skill 负责发现工作流，MCP 负责共享状态。\n- **Claude Code：** 在 `CLAUDE.md` 放同一份项目契约；用子 Agent 审就绪任务，并在最终交稿前执行\n  覆盖验证。\n- **OpenCode：** 可直接发现现有 `.agents/skills`；再接入同一条 MCP 命令共享队列。\n- **DeepSeek Harness：** 使用仓库内 `plugins/deepseek-harness` 的 npm Bundle；它会挂载官方 DSH\n  MCP 客户端并为工作区启动本服务。\n- **Manus：** 现在可直接从 GitHub 导入 Skill。需要 MCP 时，在自有 HTTPS 与 Bearer 鉴权后部署\n  HTTP 模式，且该服务根目录必须包含实际项目文件。\n- **Hermes Agent：** 从 GitHub 或技能源安装 Skill，再接入 stdio 或受保护 HTTP MCP；子 Agent 可\n  分别领取已就绪任务。\n\n`integrations/project-context/AGENTS.md.example` 故意保持很短：它只指向账本与 MCP 流程，避免把\n全部规则重复塞入每个对话上下文。\n\nFile v0.15.3:docs/audit-pipeline.md\n\n# Multi-Stage Audit Pipeline\n\nLoading every audit rule at once increases coverage, but it does not guarantee deeper checking. A crowded prompt can make a model skip dimensions, mix output contracts, or spend context on repeated instructions.\n\nThe project keeps three complementary modes:\n\n| Mode | Purpose |\n| --- | --- |\n| `build --review` | Compact editing and AI-trace guidance during generation |\n| `build --deep-review` | Expanded legacy self-review; optional narrative modules remain explicit |\n| `pipeline` | Independent, single-purpose passes over the same draft |\n\n## Complete Pipeline\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --output-dir chapter-audit\n```\n\nIt writes the established broad stages for logic, character consistency, relationship\nstance, physical continuity, AI traces, number sense, and proofreading. The higher-cost\n`voice`, `register`, `capability`, `serial`, `world`, `process`, `momentum`, `salience`, `recurrence`,\n`texture`, `fidelity`, and `sources` stages stay out unless explicitly selected or detected by\n`--auto`. `preservation` is always explicit-only because it is a high-cost source-to-rewrite\ncomparison.\n\nIt also writes `00-pattern-lint.md` and JSON as a deterministic preflight. These\nfiles contain evidence locations and a transparent editing score; they do not\nclaim to identify the author.\n\nAdd `--with-stats` to write optional `00-style-stats.md` and JSON. Statistics stay\noff by default and are editing diagnostics rather than authorship evidence.\n\nRun every generated Markdown prompt in a fresh Chatbox conversation, independent API request, or model session without prior stage memory.\n\n## Dynamic Selection\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --auto `\n  --output-dir chapter-audit\n```\n\nAutomatic mode always keeps `logic`, `ai-trace`, and `proofread`. It adds:\n\n- `character` for character-action or voice cues\n- `relationship` for dialogue, hierarchy, faction, intimacy, or secrecy cues\n- `voice` for sustained attributed dialogue; supplied context also permits shorter attributed exchanges\n- `register` for dialogue plus explicit language, dialect, honorific, or register evidence\n- `capability` for supplied context plus power, skill, authority, equipment, injury, or resource constraints\n- `serial` only when prior context is supplied and the draft is narrative\n- `world` for explicit era, world-rule, technology-system, or speculative-setting cues\n- `process` for sustained consequential process or process-to-result cues\n- `momentum` only for a multi-chapter draft or repeated continuation structure\n- `salience` only for narrative drafts with at least 4,000 characters and 12 paragraphs\n- `recurrence` only for three or more chapter headings\n- `physical` for space, movement, appearance, or prop cues\n- `texture` for cinematic opening stacks, formulaic introspection, clustered imagery,\n  detail inventory, fragment runs, or show-then-gloss cues\n- `numbers` for exact numbers with units\n- `style-match` only when `--reference` or `--reference-style` explicitly activates it\n- `fidelity` only when `--original` supplies the pre-rewrite text\n- `preservation` only with explicit selection and `--original`; automatic mode skips it\n- `sources` only when factual source files and a serious document type are both present\n\nThe `voice` stage checks stable baseline, current goal, knowledge and role constraints,\naudience, response linkage, motivated register shifts, and whether pressure-bearing\nturns receive verbal, physical, silent, interrupted, or deferred uptake without\nequating occupation with personality. The generated manifest records why every stage was selected or\nskipped. Detection is a conservative text heuristic, not complete story understanding;\nexplicitly select stages for important chapters.\n\n`register` never invents an accent from nationality or region; it checks supplied\nevidence for shared language, dialect exposure, address, particles, and switch gates.\n`capability` separates permanent baseline from temporary state and requires an earned\ngate and cost for changes or surprising outcomes.\n\n## Explicit Stages\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --stage logic `\n  --stage character `\n  --stage relationship `\n  --output-dir chapter-audit\n```\n\n`--auto` and `--stage` are mutually exclusive.\n\n`--stage style-match` is rejected unless reference material or an explicit style\ndirection is supplied. Only that stage receives the reference text.\n\n`--stage serial` is rejected unless `--context` supplies prior chapters or a\ncontinuity ledger.\n\n`--stage capability` also requires `--context`; without prior state there is no\nreliable baseline for power, skill, authority, injury, equipment, or resources.\n\n`--stage fidelity` is rejected unless `--original` supplies the pre-rewrite text.\nOnly that stage receives the original; it checks semantic preservation rather than\nstyle imitation.\n\n`--stage preservation` also requires `--original`. It compares useful ambiguity,\nintentional repetition, motifs, hesitation, subtext, speaker markers, and unresolved\npressure without turning every rough edge into a defect.\n\n`--stage sources` is rejected unless `--source` supplies factual evidence and the\ndraft is academic, news, legal, or technical. Source files enter only that stage;\nthey do not become style references or fiction context.\n\n## Recommended Order\n\n```text\npattern lint -> optional stats -> logic -> character/relationship/voice/register/capability/serial/world/process/momentum -> salience/recurrence -> physical -> AI trace/texture -> style match/fidelity/optional preservation -> numbers -> sources -> proofreading\n```\n\nAfter structural changes, re-run affected downstream stages.\n\n`pipeline` writes prompt files only. It does not call Chatbox or another model and does not merge model reports, which keeps it portable across desktop tools, local models, and API workflows.\n\nFile v0.15.3:docs/audit-pipeline.zh-CN.md\n\n# 多阶段审稿流水线\n\n一次加载所有审查规则能扩大覆盖面，但不保证每一项都检查得更深。规则过多时，模型可能漏项、混用标准，或者把大量上下文花在重复说明上。\n\n本项目同时保留三种方式：\n\n| 方式 | 用途 |\n| --- | --- |\n| `build --review` | 正文生成时的精简编辑与 AI 痕迹提醒 |\n| `build --deep-review` | 扩展版传统自审；新增高成本叙事模块仍需按需选择 |\n| `pipeline` | 把同一稿件拆成多个互不干扰的独立审稿阶段 |\n\n## 完整流水线\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --output-dir chapter-audit\n```\n\n默认生成：\n\n1. `logic`：逻辑、时间线、知识来源、动机、规则、资源和后果\n2. `character`：人物目标、声音、能力、边界、知识和变化桥梁\n3. `relationship`：谁在谁面前提谁、阵营、等级、秘密和信息权限\n4. `physical`：位置、容量、触达、服装、道具和伤势\n5. `ai-trace`：套话、公式结构、段落无推进和其他 AI 痕迹\n6. `numbers`：动作与情绪中的假精确数字\n7. `proofread`：错别字、标点、称谓、排版和机械错误\n\n为节约上下文，`voice`、`register`、`capability`、`serial`、`world`、`process`、`momentum`、`salience`、\n`recurrence`、`texture`、`fidelity` 和 `sources` 不属于默认完整流水线。只有显式选择或使用\n`--auto` 命中条件时才会生成。`preservation` 属于高成本原文对照，始终只接受显式选择。\n\n输出目录里的每个 Markdown 都是一份完整但单一职责的提示词。应在新的 Chatbox 会话、独立 API 请求或没有上一阶段聊天记忆的模型会话中分别运行。\n\n目录还会生成确定性预检 `00-pattern-lint.md` 和 JSON，包含命中位置和透明分数，\n但不据此判断作者身份。\n\n只有传入 `--with-stats` 才会额外生成 `00-style-stats.md` 和 JSON。统计默认关闭，\n只作编辑诊断，不作作者身份判断。\n\n## 动态按需加载\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --auto `\n  --output-dir chapter-audit\n```\n\n自动模式始终保留：\n\n- `logic`\n- `ai-trace`\n- `proofread`\n\n然后根据本章内容决定是否追加：\n\n- 有人物行为或声音线索：`character`\n- 有对话、等级、阵营、亲密或秘密线索：`relationship`\n- 有持续多轮对白和说话归属线索：`voice`；提供人物账本时，较短的归属明确对白也可触发\n- 对白与明确语言、方言、敬语或语域证据同时存在：`register`\n- 提供账本且出现战力、技能、权限、装备、伤势或资源约束：`capability`\n- 已提供前章/账本且正文具有连载叙事线索：`serial`\n- 有明确时代、世界规则、技术体系或架空设定：`world`\n- 同时出现关键过程和结果线索，或多次出现过程线索：`process`\n- 正文含两个以上章节标题或重复续篇标记：`momentum`\n- 至少 4000 字符、12 个段落且属于叙事长稿：`salience`\n- 至少三个章节标题：`recurrence`\n- 有位置、移动、服装、道具或空间线索：`physical`\n- 有电影式开场堆料、公式化内心解释、密集比喻、资料倾倒、连续短段或展示后再解释：`texture`\n- 有带单位的精确数字：`numbers`\n- 明确传入 `--reference` 或 `--reference-style`：`style-match`\n- 明确传入改写前原文 `--original`：`fidelity`\n- 显式选择并传入改写前原文：`preservation`；自动模式不会加入\n- 严肃文体与一个以上 `--source` 同时存在：`sources`\n\n`voice` 会对照人物稳定语言基线、当场目的、知识与角色约束、听众、回应衔接和\n变调依据，并检查关键台词或动作是否得到语言、动作、可读沉默、明确打断或延迟\n回应，不会把职业直接等同于口吻。`README.md` 清单会记录每个阶段为什么被\n选择或跳过。自动判断是保守的文本启发式，不理解完整剧情；重要章节应显式指定阶段。\n\n`register` 不根据国籍或地域凭空生成口音，只审查已经提供证据的共同语言、方言经历、\n敬语称谓、语气词和变调依据。`capability` 区分永久基线与临时伤势、装备、资源、\n冷却、克制和权限，并要求变化存在过程门与代价。\n\n## 显式选择\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --stage logic `\n  --stage character `\n  --stage relationship `\n  --output-dir chapter-audit\n```\n\n`--auto` 和 `--stage` 不能同时使用。\n\n显式选择 `--stage style-match` 时必须同时提供参考资料或明确文风方向。只有这个\n阶段会收到参考原文，其他专项审查不会被范文内容干扰。\n\n显式选择 `--stage serial` 时必须同时提供 `--context`，否则命令会拒绝运行，防止\n模型凭空补造前情。\n\n显式选择 `--stage capability` 同样必须提供 `--context`，否则没有可靠的既有战力、\n技能、权限、伤势或资源状态可供前后比较。\n\n显式选择 `--stage fidelity` 时必须同时提供 `--original`。只有这个阶段会收到改写\n前原文，它检查语义是否保留，不把原文自动当作文风范本。\n\n显式选择 `--stage preservation` 同样必须提供 `--original`。它对照有意义的含混、\n重复、母题、迟疑、潜台词、人物口吻和未结压力，避免把所有毛边都误判成错误。\n\n显式选择 `--stage sources` 时必须同时提供 `--source`，并将 `--document-type`\n设为论文、新闻、法律或技术文档，或者让自动识别获得足够的严肃文体证据。来源\n文件只进入 `sources` 阶段，不进入小说审稿和文风对齐阶段。\n\n## 推荐执行顺序\n\n先改结构，后改文字：\n\n```text\n确定性扫描 -> 可选统计 -> 逻辑 -> 人物/关系/声音/语域/能力/前情/世界/过程/推进 -> 注意力/跨章结构 -> 物理 -> AI 痕迹/文字质地 -> 文风对齐/语义保真 -> 数字 -> 来源 -> 校对\n```\n\n逻辑或剧情结构发生重写后，应重新运行受影响的后续阶段。不要先校对一段随后会被整体删除的文字。\n\n## 边界\n\n`pipeline` 负责生成分阶段提示词文件，不会主动调用 Chatbox 或任何模型，也不会自动合并模型输出。这样可以兼容 Chatbox、Codex、ChatGPT、Claude、本地模型和 API 工作流。\n\nFile v0.15.3:docs/chatbox.md\n\n# Using Advanced Human Writing & AI Humanizer in Chatbox\n\nThis guide explains how to use Advanced Human Writing & AI Humanizer with Chatbox while reducing context loss during long writing sessions.\n\nChatbox can use this project because the project outputs plain text prompt packs. You do not need a plugin integration. Generate a prompt pack with the CLI, then paste it into Chatbox as a system prompt or as the first message of a new conversation.\n\n## Why Context Can Still Be Lost\n\nChatbox context contains three parts:\n\n- system prompt\n- chat history\n- current question\n\nIn long conversations, older messages may fall outside the model's context window or be excluded by Chatbox's maximum context message setting. This project handles that by keeping a compact continuity ledger that can be pasted back into the active context.\n\n## Recommended Chatbox Setup\n\n1. Choose a model with a large context window.\n2. Increase Chatbox's maximum context messages for long writing sessions.\n3. Create a new conversation for each project, book, article series, or major chapter.\n4. Put the compiled Advanced Human Writing & AI Humanizer prompt pack in the system prompt when possible.\n5. If the UI you use does not expose a system prompt field, paste the prompt pack as the first message and ask the model to treat it as standing instructions.\n6. Keep a separate continuity ledger and update it after each major output.\n\n## Generate a Chatbox Prompt Pack\n\nFiction example:\n\n```powershell\npython -m humanwriting.cli build `\n  --style fiction `\n  --module controlled-drift `\n  --module narrative-bridges `\n  --module relationship-state `\n  --module natural-measurement `\n  --module embodied-emotion `\n  --module vocal-rhythm `\n  --module cultural-anchors `\n  --review `\n  --context examples/story-ledger.md `\n  --task \"Use these as standing instructions for this Chatbox writing session. Do not draft yet. Wait for my scene task.\"\n```\n\nSelf-media example:\n\n```powershell\npython -m humanwriting.cli build `\n  --style self-media `\n  --module imperfect-prose `\n  --module cultural-anchors `\n  --module vocal-rhythm `\n  --review `\n  --context examples/article-brief.md `\n  --task \"Use these as standing instructions for this Chatbox writing session. Do not draft yet. Wait for my article task.\"\n```\n\n## No-Loss Context Routine\n\nUse this routine every 3 to 5 substantial turns, or whenever the conversation becomes long.\n\n### 1. Ask for a Ledger Update\n\nPaste this into Chatbox:\n\n```text\nUpdate the continuity ledger only. Do not continue the draft.\n\nKeep:\n- Fixed facts\n- Active threads\n- Relationship state\n- Current audience, possible overhearers, and public/private setting\n- Public stance, private stance, and information permissions\n- Who may mention whom, allowed tone, forbidden leaks, and exception motives\n- Voice anchors\n- Stable speaker baseline, knowledge/role constraints, disclosure habits, and audience shifts\n- Current dialogue contract: purpose, desired outcomes, protected information, response tactics, and intended state change\n- Open interaction debt: consequential lines or actions awaiting uptake, refusal, interruption, consequence, or delayed payoff\n- Spatial positions and seat/standing locations\n- Physical resource modes, capacity, and occupants\n- Clothing, shoes, injuries, and prop state\n- Movement transitions already shown on-page\n- Transformation gates already shown on-page, such as folded seats, cleared tables, or opened beds\n- Current scene or section state\n- Beat bridge: previous residue, entry pressure, micro-turn, exit pressure\n- Newly true facts from the latest output\n- Newly exposed secrets, suspicions, alliance changes, and mention-policy changes\n- Open questions and unresolved pressure\n\nRemove:\n- repeated wording\n- discarded options\n- generic summaries\n- anything not needed for future continuity\n```\n\n### 2. Save the Updated Ledger\n\nCopy the ledger into a local Markdown file, for example:\n\n```text\nmy-novel-ledger.md\n```\n\n### 3. Rebuild the Prompt Pack\n\n```powershell\npython -m humanwriting.cli build `\n  --style fiction `\n  --module controlled-drift `\n  --module narrative-bridges `\n  --module relationship-state `\n  --module natural-measurement `\n  --module embodied-emotion `\n  --module vocal-rhythm `\n  --review `\n  --context my-novel-ledger.md `\n  --task \"Continue from the current scene state. Preserve all fixed facts and unresolved threads.\"\n```\n\n### 4. Start a Fresh Chatbox Thread When Needed\n\nStart a new Chatbox conversation when:\n\n- the thread exceeds 20 to 30 rounds\n- the model repeats itself\n- old facts start drifting\n- you change chapter, article, or writing goal\n\nPaste the rebuilt prompt pack at the start of the new conversation.\n\n## Auditing an Existing Draft in Chatbox\n\nWhen you need to review an already-written chapter, generate an audit pack:\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md\n```\n\nPaste the output into Chatbox. It instructs the model to extract evidence first, then check seats, barriers, reach/contact, clothing, shoes, props, and injuries.\n\nFor false precision review:\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft my-chapter.md `\n  --profile numbers\n```\n\nIt asks the model to list each exact number and decide whether to keep, soften, or delete it.\n\n### Running the Multi-Stage Pipeline in Chatbox\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --auto `\n  --output-dir chapter-audit\n```\n\nOpen a fresh Chatbox conversation for every generated stage and paste that stage's Markdown prompt. Do not run all stages in one conversation, because earlier judgments can bias later passes. Reconcile duplicate findings and repair order only after every stage finishes.\n\n## Copyable Chatbox Opening Message\n\n```text\nTreat the following as standing instructions for this writing session.\n\nRules:\n- Keep the continuity ledger active.\n- Before every draft, check fixed facts, active threads, voice anchors, and current state.\n- Before dialogue, check speaker -> listener/audience -> referenced party together with\n  baseline voice, current goal, knowledge/role constraints, response linkage, mention\n  policy, and information permissions; do not substitute occupation for personality.\n- Before shifting away from a consequential line or action, show verbal, physical,\n  silently legible, interrupted, or deliberately deferred uptake without forcing every\n  sentence into mechanical ping-pong.\n- In cars, rooms, elevators, dining areas, beds, stools, aircraft, motorcycles, or other physical spaces, check positions, resource capacity, transformations, reach, clothing, props, and movement gates before drafting.\n- Do not overwrite established facts for convenience.\n- If context is missing, make the smallest possible assumption and mark it.\n- After each long draft, output a short \"Ledger Update\" section only when I ask for it.\n- If I type \"更新账本\", update the ledger instead of continuing the draft.\n\nI will now paste the compiled Advanced Human Writing & AI Humanizer prompt pack.\n```\n\n## Practical Limits\n\nThis workflow reduces context loss, but it cannot make a model remember infinite text. For long novels, serialized fiction, research reports, or article series, the continuity ledger is the source of truth.\n\nWhen the model forgets, do not argue with the chat history. Rebuild the prompt pack from the latest ledger and continue from there.\n\nFile v0.15.3:docs/chatbox.zh-CN.md\n\n# 在 Chatbox 中使用 Advanced Human Writing & AI Humanizer\n\n这份说明用于解决两个问题：\n\n- 这个项目能不能在 Chatbox 里用\n- 长篇写作时怎样尽量不丢上下文\n\n结论：可以用。这个项目输出的是纯文本指令包，不依赖插件。你用 CLI 生成指令包后，把它粘贴到 Chatbox 的 system prompt，或者作为新会话的第一条消息即可。\n\n## 为什么还是可能丢上下文\n\nChatbox 的上下文通常包含：\n\n- System Prompt\n- 聊天历史\n- 当前问题\n\n但是模型本身有上下文窗口限制，Chatbox 也有最大上下文消息数设置。长对话写久以后，早期内容可能不再进入模型本次请求。因此，长文本不要只依赖聊天记录，要维护一份独立的 continuity ledger，也就是“连续性账本”。\n\n## 推荐设置\n\n1. 选择上下文窗口更大的模型。\n2. 在 Chatbox 设置里提高 maximum context messages / 最大上下文消息数。\n3. 每个小说、文章系列、章节或项目单独开一个会话。\n4. 能设置 system prompt 的话，把本项目生成的指令包放进去。\n5. 如果当前 Chatbox 界面没有 system prompt，就把指令包作为新会话第一条消息，并声明“以下是本会话长期规则”。\n6. 每隔几轮更新一次连续性账本，必要时用账本开启新会话。\n\n## 生成 Chatbox 指令包\n\n小说写作示例：\n\n```powershell\npython -m humanwriting.cli build `\n  --style fiction `\n  --module controlled-drift `\n  --module narrative-bridges `\n  --module relationship-state `\n  --module natural-measurement `\n  --module embodied-emotion `\n  --module vocal-rhythm `\n  --module cultural-anchors `\n  --review `\n  --context examples/story-ledger.md `\n  --task \"把这些作为 Chatbox 写作会话的长期规则。现在不要开始正文，等待我的场景任务。\"\n```\n\n自媒体写作示例：\n\n```powershell\npython -m humanwriting.cli build `\n  --style self-media `\n  --module imperfect-prose `\n  --module cultural-anchors `\n  --module vocal-rhythm `\n  --review `\n  --context examples/article-brief.md `\n  --task \"把这些作为 Chatbox 写作会话的长期规则。现在不要开始正文，等待我的文章任务。\"\n```\n\n## 不丢上下文的使用流程\n\n每 3 到 5 次长输出，或者你感觉对话变长时，做一次账本更新。\n\n### 1. 让 Chatbox 更新账本\n\n复制这段给 Chatbox：\n\n```text\n只更新连续性账本，不要继续正文。\n\n保留：\n- 固定事实\n- 活跃线索\n- 关系状态\n- 当前在场者、可能偷听者、公开或私下场合\n- 关系公开立场、私下立场和信息权限\n- 谁能在谁面前提谁、允许语气、禁泄秘密和例外动机\n- 声音锚点\n- 人物稳定说话基线、知识/角色约束、披露习惯和面对不同听众的变化\n- 当前对话契约：谈话目的、各方所求、不可透露内容、回应策略和预期状态变化\n- 未关闭的互动欠账：待回应、拒绝、打断、产生后果或延迟回收的关键台词与动作\n- 空间位置和座位/站位\n- 物理资源的形态、容量和占用者\n- 服装、鞋子、伤口和道具状态\n- 已写出的移动过渡\n- 已写出的形态转换，例如座椅放倒、桌面清空、床铺展开\n- 当前场景或章节状态\n- 节拍桥：上一拍残留、进入压力、下一步微转折、结尾压力\n- 最近输出后新增为真的事实\n- 新暴露的秘密、新产生的怀疑、阵营变化和提及策略变化\n- 未解决问题和悬而未决的压力\n\n删除：\n- 重复措辞\n- 已放弃的方案\n- 泛泛总结\n- 后续连续性不需要的信息\n```\n\n### 2. 保存账本\n\n把 Chatbox 输出的账本复制到本地 Markdown 文件，例如：\n\n```text\nmy-novel-ledger.md\n```\n\n### 3. 用最新账本重新生成指令包\n\n```powershell\npython -m humanwriting.cli build `\n  --style fiction `\n  --module controlled-drift `\n  --module narrative-bridges `\n  --module relationship-state `\n  --module natural-measurement `\n  --module embodied-emotion `\n  --module vocal-rhythm `\n  --review `\n  --context my-novel-ledger.md `\n  --task \"从当前场景状态继续。严格保留固定事实、活跃线索和未解决压力。\"\n```\n\n### 4. 必要时开启新会话\n\n出现下面情况时，建议新开 Chatbox 会话：\n\n- 当前会话超过 20 到 30 轮\n- 模型开始重复\n- 人物、设定、时间线开始漂移\n- 要进入新章节、新文章或新方向\n\n新会话第一条消息粘贴最新指令包即可。\n\n## 在 Chatbox 中审查已有稿件\n\n如果不是续写，而是检查已经写好的章节，先用本项目生成审稿指令：\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md\n```\n\n把输出粘贴到 Chatbox。它会要求模型先抽取证据表，再检查座位、隔板、触达、服装、鞋子、道具和伤口是否矛盾。\n\n如果要审查假精确数字：\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft my-chapter.md `\n  --profile numbers\n```\n\n它会要求模型列出每个精确数字，判断是必须保留、可保留、应弱化还是应删除。\n\n### 在 Chatbox 中运行多阶段审稿\n\n```powershell\nhuman-writing-skills pipeline `\n  --draft my-chapter.md `\n  --context my-novel-ledger.md `\n  --auto `\n  --output-dir chapter-audit\n```\n\n为输出目录中的每个阶段分别新建 Chatbox 会话，再粘贴对应 Markdown。不要在同一个会话里连续运行全部阶段，否则上一阶段的判断会影响下一阶段。全部完成后，再根据清单合并重复问题和修复顺序。\n\n## 可直接复制的 Chatbox 开场消息\n\n```text\n请把下面内容作为本次写作会话的长期规则。\n\n规则：\n- 始终维护连续性账本。\n- 每次写正文前，先检查固定事实、活跃线索、声音锚点和当前状态。\n- 对话前检查“说话人 -> 听话人/在场观众 -> 被提及第三方”，并核对人物基线、\n  当场目的、知识/角色约束、回应关系、提及策略和信息权限；职业不能直接替代人物性格。\n- 关键台词或动作之后，转场前必须出现语言、动作、可读沉默、明确打断或延迟回应；\n  不要让互动刺激无痕消失，也不要把每句话机械写成一问一答。\n- 涉及车内、房间、电梯、餐桌、床、板凳、飞机、摩托等空间时，先检查位置、物理资源容量、形态转换、可触达范围、服装道具和移动过渡。\n- 不要为了方便改写已确定设定。\n- 如果上下文缺失，只做最小假设，并标注为假设。\n- 长输出后，只有当我要求“更新账本”时，才输出 Ledger Update。\n- 如果我输入“更新账本”，你只更新账本，不要继续正文。\n\n接下来我会粘贴 Advanced Human Writing & AI Humanizer 生成的指令包。\n```\n\n## 现实边界\n\n这个流程能显著减少上下文丢失，但不能让模型无限记忆。长篇小说、系列文章、科研报告、世界观设定集，都应该把 continuity ledger 当成唯一可信的连续性来源。\n\n模型一旦忘了，不要和聊天记录纠缠。用最新账本重新生成指令包，然后继续。\n\nFile v0.15.3:docs/editing-tools.md\n\n# Fidelity, Statistics, and Conservative Fixes\n\nThese tools are optional. They are deliberately separated so ordinary generation\ndoes not load rewrite comparison rules, statistics, or mechanical cleanup logic.\n\n## Three Input Channels\n\n| Input | Authority | Activation |\n| --- | --- | --- |\n| `--original` | Meaning of text being rewritten | Only when supplied; enables `rewrite-fidelity` |\n| `--reference` | Transferable style features | Only with an explicit reference or style request |\n| `--source` | Factual evidence | Only for serious academic, formal, news, legal, or technical work |\n\nDo not use `--reference` as evidence for facts. Do not imitate the style of\n`--original` unless it is also deliberately supplied as a reference.\n\n## Meaning-Preserving Rewrite\n\n```powershell\nhuman-writing-skills build --style self-media --original original.md --task \"Rewrite for clarity without adding facts.\"\nhuman-writing-skills audit --draft revised.md --original original.md --profile fidelity\nhuman-writing-skills pipeline --draft revised.md --original original.md --auto --output-dir audit\n```\n\nThe fidelity pass compares entities, numbers, polarity, uncertainty, chronology,\ncausality, comparison axes, attribution, and constraints. It reports omitted,\nbroadened, reversed, reattributed, reordered, or invented meaning. It does not ask\nthe model to preserve awkward wording.\n\n## Quick Humanize And Voice Preservation\n\n```powershell\nhuman-writing-skills humanize --draft original.md --style fiction --mode quick\nhuman-writing-skills humanize --draft original.md --style fiction --mode deep\nhuman-writing-skills audit --draft revised.md --original original.md --profile preservation\n```\n\n`quick` loads surface-pattern review, rewrite fidelity, and\n`voice-ambiguity-preservation`. `deep` additionally loads high-cost cliche,\nformulaic-structure, prose-progress, imperfect-prose, and editor-loop guidance.\nExamples remain absent unless `--with-examples` is passed.\n\nThe preservation audit compares the original with the rewrite for useful ambiguity,\nintentional repetition, motifs, hesitation, subtext, speaker markers, and unresolved\ninteraction pressure. It requires `--original`, stays out of automatic pipelines,\nand must distinguish those features from unclear reference or missing grammar.\n\n## Optional Style Statistics\n\n```powershell\nhuman-writing-skills stats --draft article.md --style self-media\nhuman-writing-skills pipeline --draft article.md --auto --with-stats --output-dir audit\n```\n\nThe report includes sentence and paragraph length variation, moving-average\ntype-token ratio (MATTR), repeated trigram ratio, and explicit-transition density.\nFor Chinese and Japanese, MATTR uses character-oriented Han/kana tokens; Latin-script\nand Arabic profiles use Unicode word-like tokens. Values from different tokenization\nfamilies must not be compared directly. Short samples are marked low confidence. Raw\ntype-token ratio is intentionally omitted because it changes sharply with sample length.\n\nThese metrics are editing diagnostics, not evidence of AI authorship. Compare a\ndraft with its own genre, language, and intended voice rather than a universal\nthreshold.\n\n## Conservative Fix Preview\n\n```powershell\nhuman-writing-skills fix --draft article.md --preview\nhuman-writing-skills fix --draft article.md --output cleaned.md\nhuman-writing-skills fix --draft article.md --apply\n```\n\nPreview is the default. Automatic edits are limited to high-confidence mechanical\nresidue such as finished-prose chatbot closings and a few meaning-equivalent filler\nor wordiness reductions. The command never automatically rewrites claims, numbers,\ncomparisons, character voice, or contextual syntax.\n\n## Repeated Comparison Ladders\n\nThe linter treats two valid Chinese `比` clauses as a review signal (`STR002`). Three\nor more in one sentence become high risk (`STR004`) in fiction, webnovels,\nself-media, and general prose. Necessary data comparisons in news and academic\nwriting are suppressed. The repair is contextual: keep distinct facts with a clear\ncomparison axis; rewrite decorative escalation into one observation, action, image,\nor consequence. `fix` does not delete these structures automatically.\n\nArchive v0.15.2: 151 files, 333927 bytes\n\nFiles: agents/openai.yaml (435b), CONTRIBUTING.md (1300b), docs/agent-orchestration.md (5277b), docs/agent-orchestration.zh-CN.md (4819b), docs/audit-pipeline.md (6033b), docs/audit-pipeline.zh-CN.md (6504b), docs/chatbox.md (7460b), docs/chatbox.zh-CN.md (7173b), docs/editing-tools.md (4202b), docs/editing-tools.zh-CN.md (3869b), docs/forum-complaint-research.md (4820b), docs/forum-complaint-research.zh-CN.md (4977b), docs/ledger-extraction.md (1396b), docs/ledger-extraction.zh-CN.md (1290b), docs/long-form-consistency.md (5240b), docs/long-form-consistency.zh-CN.md (6268b), docs/number-sense.md (2079b), docs/number-sense.zh-CN.md (2336b), docs/pattern-linter.md (5135b), docs/pattern-linter.zh-CN.md (4942b), docs/physical-continuity.md (3075b), docs/physical-continuity.zh-CN.md (3753b), docs/protected-content.md (2004b), docs/protected-content.zh-CN.md (1984b), docs/reference-style.md (1859b), docs/reference-style.zh-CN.md (1802b), docs/relationship-stance-continuity.md (3247b), docs/relationship-stance-continuity.zh-CN.md (3333b), examples/article-brief.md (622b), examples/capacity-conflict-draft.zh-CN.md (628b), examples/capacity-ledger-template.md (1103b), examples/chatbox-ledger-template.md (3620b), examples/deep-fiction-task.md (1465b), examples/false-precision-draft.zh-CN.md (706b), examples/problem-car-scene-draft.md (835b), examples/problem-car-scene-draft.zh-CN.md (683b), examples/reference-style-draft.zh-CN.md (184b), examples/reference-style-source.zh-CN.md (371b), examples/relationship-stance-ledger.zh-CN.md (1992b), examples/story-ledger.md (4053b), examples/vehicle-scene-ledger.md (3995b), humanwriting/__init__.py (127b), humanwriting/audit_queue.py (9005b), humanwriting/cli.py (28803b), humanwriting/compiler.py (36582b), humanwriting/config.py (3848b), humanwriting/detection.py (18713b), humanwriting/fixer.py (4556b), humanwriting/ledger.py (2446b), humanwriting/linter.py (61773b), humanwriting/longform.py (32925b), humanwriting/mcp_server.py (24993b), humanwriting/original.py (1341b), humanwriting/pipeline.py (13032b), humanwriting/precommit.py (1105b), humanwriting/protection.py (10445b), humanwriting/reference.py (3700b), humanwriting/skills.py (2090b), humanwriting/source.py (1814b), humanwriting/statistics.py (9087b), integrations/mcp/mcp.json.example (202b), integrations/project-context/AGENTS.md.example (741b), LICENSE (1107b), marketplaces/skillhub-overview.zh-CN.md (5791b), plugins/deepseek-harness/cordis.patch.yml (351b), plugins/deepseek-harness/index.js (840b), plugins/deepseek-harness/LICENSE (1092b), plugins/deepseek-harness/package.json (1158b), plugins/deepseek-harness/README.md (1278b), plugins/deepseek-harness/README.zh-CN.md (1005b), pyproject.toml (1378b), README.es.md (4863b), README.fr.md (4806b), README.md (45725b), README.pt-BR.md (4495b), README.zh-CN.md (45978b), scripts/build_skillhub_package.py (4399b), skill-card.md (2867b), SKILL.md (9053b), skills/__init__.py (77b)\n\nFile v0.15.2:SKILL.md\n\n---\nname: human-writing-skills\ndescription: Advanced multilingual AI humanizer for natural rewriting, fiction editing, long-form audit and continuity, verified chunked agent review, translationese review, and character consistency. Humanize AI text, remove robotic tone, edit fiction and novels, continue webnovel chapters, proofread writing, and audit story continuity, character voice, dialogue register and performance, scene geography, relationships, numbers, citations, source meaning, and translated-text fidelity. Use for AI writing cleanup, supplied-sample style matching, long-context fiction, essays, news, official, academic, legal, and technical prose. Trigger on humanize AI text, de-AI writing, natural rewriting, novel writing assistant, story consistency checker, scene ending audit, reflective ending, chunked audit, long-form agent audit, translationese audit, style consistency review, character consistency audit, dialogue audit, dialogue action audit, 增强版去 AI 写作 Skill、高级 AI 写作工具、去AI味、去AI写作、消除AI腔、AI人性化改写、AI文本润色、AI文章润色、小说润色、小说续写、AI式结尾、生硬结尾审查、无意义升华、长文一致性、长篇审查、分块审查、文风统一、统一文风、人物设定统一、人物一致性审查、跨章一致性、小说审查、报告审查、人物口吻、人物对白审查、对话生硬、对白动作、方言语域、翻译腔、翻译审查、战力设定、场景空间审查.\n---\n\n# Advanced Human Writing & AI Humanizer\n\nHumanize AI-shaped text, write or continue genre-aware prose, and audit long-form\ncontinuity without flattening a specific voice. Use the smallest set of modules\nthat covers the task. Keep project facts, prior chapters, rewrite originals, and\ncontinuity ledgers separate from optional style references.\n\n## Capability Layers\n\n- The `humanwriting/` Python package is executable. Its CLI deterministically\n  compiles prompts, locates recurring text patterns, calculates diagnostics,\n  previews conservative fixes, checks protected content, and writes staged audits.\n- The Markdown files under `skills/` are model-executed writing and editorial\n  modules. They are selected by the compiler and are intentionally not pretending\n  to be deterministic NLP algorithms.\n- A normal installation includes both layers. Verify the executable layer with\n  `human-writing-skills list --kind module` and run a real draft through `lint`,\n  `fix`, `verify`, or `pipeline` rather than judging the package from `SKILL.md` alone.\n- `human-writing-mcp` is an optional project-local coordination layer. It gives\n  separate agents bounded long-form assignments, stores receipts, and refuses\n  reconciliation until coverage verification passes. It does not call a model.\n\n## Quick Humanize Route\n\n- For a supplied draft, use `humanize --mode quick` for surface patterns,\n  rewrite fidelity, and voice/ambiguity preservation.\n- Use `--mode deep` only when structural repetition, cliches, paragraph stagnation,\n  or broad editorial reconstruction needs a separate pass.\n- Load `humanize-examples` only after an explicit request or `--with-examples`.\n- Do not activate rewrite-preservation modules for unrelated new drafting.\n\n## Workflow\n\n1. Select one base style from `skills/`: `fiction`, `webnovel`, `argumentative`,\n   `news-report`, `formal-document`, `self-media`, or `academic-paper`.\n2. Read only the relevant modules. Add continuity, spatial, relationship, number,\n   dialogue, dialogue-performance, register, capability, world, process, salience, recurrence, source, rhythm, preservation,\n   or AI-trace modules when the text actually needs them.\n   `narrative-naturalness-audit` is reserved for deep or explicit AI-trace review of\n   narrative prose; it is not loaded for ordinary quick humanization or serious documents.\n   For fiction with spoken or wordless interpersonal exchanges, use\n   `dialogue-voice-audit` and `dialogue-performance-audit` (CLI: `--profile voice`).\n   Check both participants and separate speech from simultaneous action; establish\n   reception, refusal, or an evidenced pending state before assuming completion.\n   During drafting, reconsider this route when an interaction enters the scene even\n   if the original request only said to continue a chapter.\n3. Treat user facts and `--context` as authoritative. Never borrow facts from a\n   style sample. When `--original` is supplied for a rewrite, activate\n   `rewrite-fidelity` and preserve meaning without preserving awkward wording.\n   In fiction and webnovels, do not let time/place mini-headings replace scene bridges.\n4. Activate `reference-style-alignment` only when the user supplies reference\n   material, gives an explicit style direction, or directly asks to match a style.\n5. Treat `--source` as factual evidence only. Activate `source-grounding` only for\n   serious academic, formal, news, legal, or technical work with explicit source files.\n6. For important revisions, run deterministic `lint`, then independent audit\n   profiles, then `verify` protected content against the source. Run `stats` only\n   when distributional diagnostics help; use `fix` as a preview before writing.\n7. Keep `voice`, `serial`, `world`, `process`, `momentum`, `salience`, `recurrence`,\n   `ending`, `texture`, `fidelity`, `preservation`, examples, and `sources` separate from the default audit. Activate them explicitly\n   or through `pipeline --auto`. `serial` requires context, `recurrence` requires at\n   least three chapters, `fidelity` requires `--original`, `preservation` requires\n   both explicit selection and `--original`, examples require an explicit request,\n   and `sources` requires both a serious document and `--source`.\n8. For a book-length manuscript or large report, use `chunk-audit` instead of placing\n   the whole draft in one prompt. Supply `--outline` or `--context` for authoritative\n   character or report rules. Use `--agent-mode deep` only when every block needs a\n   visible coverage receipt, then run `verify-chunk-audit` before reconciliation. Use\n   `--translationese` only for an explicitly translated or localized work. Read\n   `docs/long-form-consistency.md` only for this workflow.\n9. When multiple agents or sessions share a long-form project, use\n   `human-writing-mcp` rather than copying the entire manuscript into each chat.\n   Agents must claim a task, submit a complete receipt, and pass coverage\n   verification before one final reconciliation task is released. Read\n   `docs/agent-orchestration.md` only when an MCP-capable host is available.\n\n## Commands\n\n```powershell\nhuman-writing-skills build --style fiction --context ledger.md --task \"Continue the scene.\"\nhuman-writing-skills humanize --draft chapter.md --style fiction --mode quick\nhuman-writing-skills humanize --draft article.md --style self-media --mode deep --with-examples\nhuman-writing-skills build --style fiction --reference sample.md --task \"Match the sample's restrained rhythm.\"\nhuman-writing-skills build --style self-media --original original.md --task \"Rewrite without adding facts.\"\nhuman-writing-skills audit --draft chapter.md --context ledger.md --profile physical\nhuman-writing-skills audit --draft chapter.md --profile voice\nhuman-writing-skills audit --draft chapter.md --context ledger.md --profile serial\nhuman-writing-skills audit --draft chapters.md --profile momentum\nhuman-writing-skills audit --draft chapters.md --profile recurrence\nhuman-writing-skills audit --draft chapter.md --profile process\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile ending\nhuman-writing-skills audit --draft paper.md --document-type academic-paper --source study.md --profile sources\nhuman-writing-skills audit --draft revised.md --original original.md --profile fidelity\nhuman-writing-skills audit --draft revised.md --original original.md --profile preservation\nhuman-writing-skills pipeline --draft chapter.md --context ledger.md --auto --with-stats --output-dir audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline outline.md --output-dir novel-audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline outline.md --agent-mode deep --output-dir novel-agent-audit\nhuman-writing-skills verify-chunk-audit --package-dir novel-agent-audit\nhuman-writing-skills chunk-audit --draft report-es.md --style news-report --source sources.md --translationese --agent-mode deep --output-dir report-audit\nhuman-writing-mcp --root C:\\writing-project\nhuman-writing-skills lint --draft chapter.md --style fiction\nhuman-writing-skills stats --draft chapter.md --style fiction\nhuman-writing-skills fix --draft chapter.md --preview\nhuman-writing-skills verify --source original.md --candidate revised.md\n```\n\nRead `README.md` or `README.zh-CN.md` for user-facing guidance. Read files under\n`docs/` only for the workflow being used. Do not claim detector evasion or infer\nauthorship from stylistic patterns; frame results as editing evidence.\n\nFile v0.15.2:plugins/deepseek-harness/README.md\n\n# Advanced Human Writing for DeepSeek Harness\n\nThis DeepSeek Harness bundle mounts the `human-writing-mcp` tools from the\nparent repository. It coordinates bounded long-form reviews locally: create a\nplan, let separate agents claim focused work, require complete coverage\nreceipts, and unlock the final reconciliation only after verification passes.\n\n## Prerequisites\n\nInstall the Python package in the environment selected by `PYTHON` or `python`:\n\n```powershell\npip install human-writing-skills\n```\n\nFor development from this repository:\n\n```powershell\npip install -e .\n```\n\n## Install\n\nAfter publishing this folder to npm, add it to a DeepSeek Harness profile:\n\n```powershell\ndsh plugin --profile web add dsh-advanced-human-writing\n```\n\nRestart the affected DSH profile. The bundle uses the workspace as its allowed\nfile root. Set `PYTHON` before launching DSH when the desired interpreter is\nnot named `python`.\n\nThe mounted tools are `plan_long_form_audit`, `list_audit_tasks`,\n`claim_audit_task`, `submit_audit_report`, `verify_audit_coverage`,\n`get_reconciliation_task`, and `read_project_context`.\n\nThis package does not send drafts to a third party and does not call a model by\nitself. The active DSH agent performs each assigned review and must submit its\nown report.\n\nFile v0.15.2:README.md\n\n# Advanced Human Writing & AI Humanizer\n\n> Reusable multilingual writing `SKILLS` for natural prose, genre-aware style, and long-form continuity.\n\n**Advanced AI humanizer and de-AI writing toolkit** for natural rewriting,\nAI text cleanup, fiction editing, novel continuation, chunked long-form audit,\nwriting style unification, and character consistency review.\n\n[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)\n[![Python](https://img.shields.io/badge/Python-3.9%2B-blue.svg)](pyproject.toml)\n[![Zero Dependencies](https://img.shields.io/badge/dependencies-zero-brightgreen.svg)](pyproject.toml)\n\n[中文说明](README.zh-CN.md) | English | [Español](README.es.md) | [Português (Brasil)](README.pt-BR.md) | [Français](README.fr.md)\n\nAdvanced Human Writing & AI Humanizer is an open-source, modular skill pack and lightweight prompt compiler for natural multilingual AI-assisted writing. The package, repository, and ClawHub slug remain `human-writing-skills` for compatibility.\n\nThis is not an empty prompt collection. The repository contains two deliberately\ndifferent capability layers:\n\n- **Executable Python tools:** `humanwriting/` provides the installable\n  `human-writing-skills` CLI for deterministic lint findings with evidence spans,\n  text statistics, conservative fix previews, protected-content verification,\n  prompt compilation, and staged audit-file generation.\n- **Model-executed editorial modules:** `skills/*.md` contains genre and review\n  instructions selected on demand by that compiler. These modules guide the writing\n  model; they do not falsely present subjective literary judgment as a deterministic\n  algorithm.\n\nThe test suite exercises both the executable layer and module-selection gates.\n\n## Agent Orchestration, MCP, And DeepSeek Harness Plugin\n\nFor book-length novels, report series, or research coverage that cannot be\ntrusted to one chat window, install the dependency-free `human-writing-mcp`\nserver. It makes the long-form task graph executable across Codex, Claude Code,\nOpenCode, DeepSeek Harness, Manus, and Hermes: agents claim bounded tasks,\nsubmit coverage receipts, and cannot unlock reconciliation until every required\nreview is complete. The service stays inside a chosen project root and does not\ncall a model or upload drafts.\n\n```powershell\nhuman-writing-mcp --root C:\\writing-project\n```\n\nThe repository also ships a native DeepSeek Harness npm/Cordis bundle under\n[`plugins/deepseek-harness`](plugins/deepseek-harness). It mounts the official\nDSH MCP client and starts the same verified local coordination service. See the\n[agent orchestration guide](docs/agent-orchestration.md) and\n[DeepSeek Harness plugin guide](plugins/deepseek-harness/README.md).\n\nThe MCP server also supports day-to-day single-document work: `lint_text`,\n`get_style_statistics`, `verify_protected_content`, `compile_humanize_prompt`,\n`compile_audit_prompt`, and `compile_ledger_extraction`. It advertises a small\nnative prompt menu (`humanize-quick`, `dialogue-audit`, `continuity-audit`,\n`serious-rewrite`, and more) for hosts that implement MCP Prompts. These tools\nstay local and return evidence or compiled instructions; they never send a draft\nto a model themselves.\n\n## Ledger Auto-Extraction And Project Defaults\n\nLong-form continuity should not require hand-maintaining every fact from scratch.\n`extract-ledger` compiles an evidence-first prompt for turning existing chapters\ninto a **candidate** continuity ledger. Its output distinguishes observed facts,\ninferences, conflicts, and unknowns, and requires a quote or location for every\nproposed state, object, obligation, injury, resource change, or spatial relation.\nReview it before making it canonical.\n\n```powershell\nhuman-writing-skills extract-ledger --draft chapters-01-10.md --context novel-ledger.md --output ledger-extraction-prompt.md\n```\n\nUse `.humanwriting.json` in a project root to persist only lightweight defaults:\nstyle, document type, a relative ledger path, and lint allow-list entries. Explicit\nCLI flags win, and the file cannot silently activate deep profiles or expensive\nreference/source passes. See the [ledger extraction guide](docs/ledger-extraction.md).\nPass `--no-project-config` when a one-off command must ignore the nearest project file.\n\n```json\n{\n  \"style\": \"fiction\",\n  \"document_type\": \"fiction\",\n  \"context\": \"novel-ledger.md\",\n  \"allow\": [\"END001\"]\n}\n```\n\n## Editor, CI, And Python Distribution\n\nText-facing commands accept `--draft -` for standard input. `lint --format github`\nemits GitHub Actions annotations, while `list --kind rule` prints the current rule\ncatalog with severity, category, and repair direction.\n\n```powershell\ngit diff -- docs\\ | human-writing-skills lint --draft - --style general --format github --source-name docs-change.md\nhuman-writing-skills list --kind rule --format json\n```\n\nEach GitHub Release builds a wheel and source distribution, then attaches them as\nRelease assets. PyPI publication uses GitHub Trusted Publishing only after the\nrepository variable `PYPI_PUBLISH_ENABLED=true` and the PyPI trusted publisher\nhave been configured; no PyPI token is stored in this repository.\n\nIt helps a writing agent move away from generic, template-shaped output and toward prose that has intention, texture, continuity, and genre discipline. The project is especially useful for long-form generation, where characters, settings, arguments, facts, and unresolved threads often drift after several passages.\n\nThe goal is not deception. The goal is better writing: clearer instructions, stronger revision habits, and reusable style constraints that make AI-assisted drafts feel edited by a human.\n\n## Long-Form Audit And Style Unification\n\nThe executable `chunk-audit` workflow splits a year-long novel, article series, or\nlarge report at natural boundaries. Each body span is audited once, with a small\nread-only lead-in and the same user-confirmed style baseline plus outline or project ledger.\nIt writes independent chunk prompts, deterministic cross-chunk style diagnostics, and\na reconciliation prompt for model- or prompt-version drift in narration, character\ndialogue, terminology, and section function.\n\nFor fiction, `--outline` or `--context` makes supported goals, knowledge, relationships,\nlimits, abilities, and speaker voice canonical. Without it, character inferences remain\nprovisional. News, academic, official, and report workflows instead align terminology,\nfacts, attribution, claim scope, and section purpose without loading fiction rules.\n\n```powershell\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline novel-outline.md --output-dir novel-audit\n```\n\nDiscovery terms: **long-form audit, chunked manuscript audit, writing style\nunification, style consistency review, character consistency audit, cross-chapter\ncontinuity, novel audit, and report review**. See the\n[long-form consistency guide](docs/long-form-consistency.md).\n\n## Earned Scene And Document Endings\n\nAI-assisted drafts often reach a real stopping point and then append a scenic dissolve,\nshared silence, life lesson, future-facing reflection, or summary of what the scene\nalready showed. The `earned-ending-audit` finds this **reflective bookend / false\nclosure** pattern in Chinese and English by locating the last meaningful change and\napplying a deletion test. It does not ban sunsets, silence, reflection, or lyrical prose.\n\n- Fiction and webnovels stop on an earned consequence, decision, discovery, changed\n  object, live pressure, or image whose meaning changed inside the scene.\n- Hard news ends on the last useful verified fact, response, constraint, or next step;\n  feature kickers must add meaning instead of manufacturing uplift.\n- Academic, technical, and official writing ends with supported findings, limits,\n  implications, decisions, owners, or deadlines rather than a ceremonial conclusion.\n\n```powershell\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile ending\nhuman-writing-skills lint --draft chapter.md --style fiction\n```\n\nThe full module loads only through the explicit `ending` profile, while `END001` provides a lightweight\ndeterministic preflight for narrative endings. Discovery terms: **AI story ending,\nreflective ending, scene ending audit, chapter ending audit, formulaic conclusion,\nfalse closure, AI reflective bookend, and can't-help-but-reflect ending**.\n\n## Repetition, Exposition, And Scene Economy\n\nSome drafts avoid obvious stock phrases yet still feel generated because they repeat a\nline, image, action sequence, or narrator explanation with no changed consequence.\nThe narrative-only `repetition-exposition-audit` separates intentional refrains and\nneeded recap from exact echoes, repeated choreography, inventory-like viewpoint scans,\nand explanation that merely tells the reader what dialogue or action already showed.\n\n```powershell\n# Isolated, token-bounded narrative pass\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile repetition\n\n# Long drafts: add it only when matching narrative cues appear\nhuman-writing-skills pipeline --draft chapter.md --lint-style fiction --auto --output-dir chapter-audit\n```\n\n`REP001` flags a non-trivial verbatim sentence echo; `NAT005` and `NAT006` are\nlow-severity density prompts for repeated interpretive narration and action frames.\nThey are editing evidence, not authorship claims, and can be allowlisted. Quick\nhumanization and serious-document workflows do not load this module.\n\n## Why This Exists\n\nAI writing often fails in predictable ways:\n\n| Problem | What this project adds |\n| --- | --- |\n| Generic \"AI voice\" | Concrete revision checks for rhythm, specificity, and empty phrasing |\n| Repeated not-X/is-Y, is-X/not-Y, or chained Chinese 比 frames | Family- and density-based checks that preserve necessary correction and real comparison |\n| Rewriting silently changes facts, uncertainty, or causal meaning | An opt-in original-text fidelity pass with a claim ledger and invention checks |\n| Humanizing washes out hesitation, motifs, subtext, or speaker identity | A source-backed preservation ledger that separates useful ambiguity from real defects |\n| Fluent-looking sentences drop a word, object, or connector clause | A separate final pass over predicate slots, parallel structure, and references |\n| Surface AI patterns recur across a passage | Genre-aware checks for vague attribution, inflated significance, false ranges, synonym cycling, formatting habits, and comparison ladders |\n| Fiction is chopped up by time/place mini-headings | Narrative-only checks that preserve titles and chapters but require scene changes to move through prose |\n| A finished scene grows a scenic, reflective, or moralizing tail | Last-meaningful-change and deletion tests for false closure, with genre-specific ending contracts |\n| A scene replays the same action, reaction, or explanation until it becomes mechanical | Narrative-only evidence for exact echoes, repeated action frames, viewpoint inventory, and redundant gloss |\n| One style fits every genre | Separate Markdown `SKILLS` for different writing forms |\n| Long text loses continuity | A compact ledger for facts, plot, promises, and voice anchors |\n| Prose and character dialogue drift across months or model versions | A fixed baseline, canonical outline, unique audit chunks, and cross-chunk reconciliation |\n| Dialogue sounds interchangeable or out of character | Generation and review against baseline voice, scene goal, knowledge, audience, and pressure |\n| Dialogue ends in a stock gesture or scenic gloss instead of a real exchange | A dialogue-performance pass checks listener uptake, purposeful physical beats, and the changed option or carried debt |\n| Dialect, honorifics, particles, or foreign language jump between characters | An evidence-backed language-identity card with motivated switch gates |\n| A consequential line or action receives no uptake before the prose cuts away | Response-obligation checks and deferred interaction debt |\n| Power, skill, authority, equipment, injury, or resources drift | Permanent/temporary state separation and earned transition gates |\n| Prompts become messy | A CLI that compiles style, context, and task into one clean instruction pack |\n| Advice stays abstract | Rules are written as observable editing actions |\n\n## Built-In Style Skills\n\n| Skill | Use it for | Main focus |\n| --- | --- | --- |\n| `fiction` | literary or commercial fiction | point of view, scene pressure, character behavior |\n| `argumentative` | essays and opinion pieces | thesis, evidence, counterargument, logical flow |\n| `news-report` | news-style reports | factual order, attribution, neutral wording |\n| `self-media` | social posts and creator essays | useful voice without empty hype |\n| `academic-paper` | research writing | cautious claims, structure, terminology |\n| `formal-document` | official and administrative documents | authority, scope, responsibility, action, deadline, restrained register |\n| `webnovel` | serialized genre fiction | hooks, payoffs, power rules, continuity |\n\n## Deep Human-Trace Modules\n\nThese modules target deeper AI-writing artifacts, not only surface phrases.\n\n| Module | What it repairs |\n| --- | --- |\n| `controlled-drift` | overly smooth logic, no associative movement, no unfinished thought |\n| `narrative-bridges` | weak scene turns, generic transitions, paragraphs that do not cause each other |\n| `relationship-state` | relationships that reset, dialogue without leverage, forgotten secrets or boundaries |\n| `relationship-stance-audit` | audience-specific stance checks for rivalries, affairs, factions, hierarchy, sects, and family politics |\n| `logic-causality-audit` | cause, timeline, knowledge, motive, rule, resource, and consequence failures |\n| `character-consistency-audit` | character goal, voice, competence, boundary, knowledge, and change-gate drift |\n| `dialogue-voice-audit` | character-fit dialogue plus verbal, physical, silent, interrupted, or deferred uptake for consequential turns |\n| `dialogue-performance-audit` | uses the voice audit's response map to test whether a physical beat changes the exchange rather than decorating it |\n| `speech-register-continuity` | evidence-backed language, dialect exposure, honorifics, particles, address, and switch gates |\n| `capability-state-audit` | power, skill, authority, equipment, injury, resources, cooldowns, counters, and transitions |\n| `serial-reentry` | recap dumps and chapter resets when prior chapters or a ledger are supplied |\n| `long-form-style-consistency` | chunked long-form style, character-setting, and speaker-voice reconciliation |\n| `chapter-momentum-audit` | atmosphere-only chapters, missing payoffs, discarded residue, and unsupported hooks |\n| `world-ontology-audit` | incompatible era, technology, institution, social practice, or speculative rule |\n| `process-earnedness-audit` | promised processes skipped before an unsupported result |\n| `attention-budget-audit` | low-value expansion and semantic echoes displacing consequential material |\n| `chapter-pattern-audit` | repeated chapter architecture across three or more chapters |\n| `narrative-distance-control` | unmotivated zoom, missing orientation, and viewpoint-distance drift |\n| `imagery-load-audit` | stacked comparisons, competing sensory channels, and show-then-gloss repetition |\n| `paragraph-rhythm-audit` | mechanical one-line paragraph runs and overloaded long blocks |\n| `detail-disclosure-audit` | biography and appearance inventories delivered before the scene uses them |\n| `scene-entry-audit` | exact-time/location/weather/outfit opening bundles before pressure-bearing action |\n| `natural-measurement` | false precision: tiny exact measures and counted micro-actions in narrative prose |\n| `cliche-phrase-audit` | stock phrases, generic body cues, empty emotion labels, and dead transitions |\n| `formulaic-structure-audit` | triplets, bidirectional contrast frames, chained comparisons, and overly neat closure |\n| `prose-progress-audit` | static paragraphs and pressure-bearing interactions abandoned before uptake or explicit deferral |\n| `narrative-naturalness-audit` | in deep or explicit AI-trace review, catches recurring six-beat scene recipes and copied entry/closure cadence |\n| `repetition-exposition-audit` | narrative-only exact echoes, repeated action choreography, inventory drift, and redundant explanatory narration |\n| `earned-ending-audit` | reflective bookends, scenic dissolves, false closure, stock kickers, and conclusions added after the last meaningful change |\n| `imperfect-prose` | prose that is too clean, too symmetrical, or too polished |\n| `vocal-rhythm` | flat cadence and missing read-aloud breath points |\n| `embodied-emotion` | emotion labels without body, action, contradiction, or perception |\n| `cultural-anchors` | vacuum prose with no era, place, community, or material detail |\n| `spatial-blocking` | character teleportation and confused front/back/left/right blocking |\n| `occupancy-capacity` | over-occupied or mode-ambiguous seats, benches, beds, stools, aisles, and surfaces |\n| `appearance-prop-continuity` | clothing, shoes, props, injuries, and daily-detail drift |\n| `physical-continuity-audit` | optional light manual checklist; do not combine with the forensic physical profile |\n| `proofreading-audit` | final omissions, predicate slots, stranded connectors, references, punctuation, naming, and layout |\n| `style-matrix` | the mistake of applying one generic \"human voice\" to every genre |\n| `editor-loop` | one-shot drafting without a critical human-editor pass |\n| `ai-trace-rubric` | vague feedback like \"sounds AI\" without diagnosis |\n| `reference-style-alignment` | explicit reference material into transferable voice features without copying content |\n| `rewrite-fidelity` | meaning drift, invented specificity, reversed polarity, and altered uncertainty when an original is supplied |\n| `voice-ambiguity-preservation` | over-clean rewrites that erase useful ambiguity, repetition, motifs, hesitation, subtext, or speaker markers |\n| `humanize-examples` | an explicit-only before/after repair library; never loaded as a source or default style sample |\n| `surface-pattern-audit` | recurrent formatting, false ranges, synonym cycling, and narrative mini-headings without global bans |\n| `protected-content` | accidental changes to numbers, citations, equations, URLs, code, quotes, and required terms |\n| `source-grounding` | claim-to-source checks for serious documents with explicit factual sources |\n\n## Quick Start\n\n```powershell\ngit clone https://github.com/whh110112/human-writing-skills.git\ncd human-writing-skills\npython -m pip install .\n\nhuman-writing-skills list --kind style\nhuman-writing-skills list --kind module\nhuman-writing-skills build --style fiction --context examples/story-ledger.md --task \"Write the next scene.\"\nhuman-writing-skills humanize --draft chapter.md --style fiction --mode quick\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline novel-outline.md --output-dir novel-audit\nhuman-writing-skills lint --draft chapter.md --style fiction\nhuman-writing-skills verify --source original.md --candidate revised.md\n```\n\nYou can also run directly from the source checkout with `python -m humanwriting.cli ...`. The `build` and `humanize` commands print instruction packs that can be pasted into Codex, ChatGPT, Claude, local LLM tools, or another writing agent.\n\n## Quick Humanize\n\n`humanize` is the low-friction rewrite route. It treats `--draft` as the original,\nkeeps the same language and genre by default, and preserves meaning before changing\nsurface style.\n\n```powershell\n# Minimum stack: surface patterns + fidelity + voice/ambiguity preservation\nhuman-writing-skills humanize --draft chapter.md --style fiction\n\n# Add structural editor passes only when the draft needs them\nhuman-writing-skills humanize --draft article.md --style self-media --mode deep\n\n# Examples remain opt-in and are never treated as factual or stylistic source material\nhuman-writing-skills humanize --draft chapter.md --style fiction --with-examples\n```\n\n`quick` does not load cliche, formulaic-structure, paragraph-progress, or editor-loop\nmodules. `deep` adds those high-cost passes. `humanize-examples` loads only with\n`--with-examples`; `voice-ambiguity-preservation` loads only for supplied-text\nhumanization or an explicit preservation audit.\n\n## Multilingual Scope\n\nThe skill instructions have no Chinese-only gate: they can guide fiction and serious\nprose in English, Japanese, French, Spanish, Portuguese, Arabic, Latin, and other\nlanguages supported by the selected model. Deterministic lexical rules are naturally\nlanguage-specific, while structural continuity and review remain language-agnostic.\nThe narrative heading scanner recognizes time cards across the languages above, and\n`stats` profiles Han, kana, Arabic, and several Latin-script language families. Use\ngenre context and human review for mixed-language or low-resource text.\n\n## Example Output Shape\n\n```text\n# Core Directive\n# Continuity Protocol\n# Selected Skill: fiction\n# Project Context\n# Task\n# Output Contract\n```\n\nThis format keeps the model focused on the current task while still carrying the previous facts, style decisions, and unresolved threads.\n\n## Explicit Reference Style\n\nReference matching is opt-in. It activates only with `--reference`,\n`--reference-style`, or explicit task wording such as \"match this voice.\" A\ncontinuity ledger by itself never activates it.\n\n```powershell\nhuman-writing-skills build `\n  --style fiction `\n  --context examples/story-ledger.md `\n  --reference examples/reference-style-source.zh-CN.md `\n  --task \"Continue the scene while matching the reference's restrained rhythm.\"\n\nhuman-writing-skills audit `\n  --draft my-chapter.md `\n  --reference examples/reference-style-source.zh-CN.md `\n  --profile style-match\n```\n\nThe compiler extracts point of view, rhythm, register, imagery, description,\ndialogue cadence, emotion handling, and transitions. Plot facts still come from\n`--context`; names, events, and distinctive phrases must not be copied from the\nreference. See [docs/reference-style.md](docs/reference-style.md).\n\n## Original-Text Fidelity\n\nUse `--original` only when revising an existing text and meaning must remain stable.\nIt activates a dedicated fidelity module for rewrite and review; ordinary drafting\ndoes not pay this token cost.\n\n```powershell\nhuman-writing-skills build `\n  --style self-media `\n  --original original.md `\n  --task \"Rewrite for clarity without adding facts or strengthening claims.\"\n\nhuman-writing-skills audit `\n  --draft revised.md `\n  --original original.md `\n  --profile fidelity\n```\n\n`--original` is semantic authority, `--reference` is style evidence, and `--source`\nis factual evidence for serious documents. They are deliberately isolated so a\nstyle sample cannot rewrite facts and an original cannot silently become a style\ntarget. See [docs/editing-tools.md](docs/editing-tools.md).\n\n## Serious-Document Sources\n\n`--source` is separate from `--reference`. It activates `source-grounding` only for\nacademic, news, legal, or technical work and builds a claim-to-source evidence map.\nFiction, webnovels, self-media, and casual answers do not auto-load it.\n\n```powershell\nhuman-writing-skills audit `\n  --draft paper.md `\n  --document-type academic-paper `\n  --source study-a.md `\n  --source study-b.md `\n  --profile sources\n```\n\nThe audit separates source existence from claim support. Without external registry\naccess, it marks citation metadata as unverified instead of inventing a verdict.\n\n## Long-Form Continuity\n\nFor longer works, this project recommends a small ledger instead of relying only on a large context window. Use context in this order: canonical ledger, latest confirmed state, recent chapters, relevant retrieved older spans, then explicitly uncertain inference. Retrieved text is recall evidence and cannot overwrite a later canonical state.\n\nThe ledger tracks:\n\n- fixed facts: names, dates, locations, relationships, rules, timeline\n- active threads: unresolved conflicts, clues, promises, open arguments\n- relationship state: who knows, wants, hides, owes, refuses, or holds leverage\n- relationship stance: public/private posture, current audience, mention policy, forbidden leaks, and exception motives\n- voice anchors: point of view, diction, directness, disclosure habits, domain limits, audience shifts, taboo phrases\n- language identity: shared scene language, demonstrated dialect/second-language exposure, address forms, particles, and switch gates\n- capability state: permanent power/skill/authority plus temporary injury, equipment, resources, cooldowns, counters, costs, and transition gates\n- dialogue contract: who speaks to whom, why now, desired listener action, protected information, and intended state change\n- interaction debt: which consequential line or action still awaits uptake, refusal, interruption, consequence, or delayed payoff\n- current state: where the previous passage ended and what must connect next\n- beat bridge: previous residue, entry pressure, micro-turn, and exit hook\n- change log: what became newly true in the latest output\n\nSee [examples/story-ledger.md](examples/story-ledger.md) for a fiction example.\n\n`speech-register-continuity` auto-loads only for fiction/webnovel dialogue when the task or ledger contains explicit language, regional, dialect, honorific, or register evidence. It can also be selected with `audit --profile register`; region or nationality never licenses an invented accent.\n\n`capability-state-audit` loads during generation only when the current task names a capability constraint. Automatic pipeline review additionally requires context, so ordinary dialogue scenes do not pay its Token cost. Select it explicitly with `audit --profile capability --context ledger.md` when needed.\n\n## Chatbox\n\nYes, this project works in Chatbox because it outputs plain text prompt packs. For long writing sessions, use the continuity ledger as the source of truth and paste the compiled prompt pack into Chatbox's system prompt or first message.\n\n- English guide: [docs/chatbox.md](docs/chatbox.md)\n- Chinese guide: [docs/chatbox.zh-CN.md](docs/chatbox.zh-CN.md)\n- Ledger template: [examples/chatbox-ledger-template.md](examples/chatbox-ledger-template.md)\n\n## Physical Continuity\n\nFor scenes where space matters, such as cars, elevators, hospital rooms, dining tables, and bedrooms, use `--strict-continuity`. It adds occupancy, spatial blocking, and appearance/prop generation guards. Use `audit --profile physical` for one evidence-first forensic pass on an existing draft; it owns capacity, blocking, appearance, props, barriers, reach, and body-state contradictions in one ledger.\n\n```powershell\npython -m humanwriting.cli build `\n  --style fiction `\n  --strict-continuity `\n  --review `\n  --context examples/vehicle-scene-ledger.md `\n  --task \"Continue the car argument. Every seat change must have an on-page transition. Keep clothing and props consistent.\"\n```\n\n- Guide: [docs/physical-continuity.md](docs/physical-continuity.md)\n- Vehicle ledger example: [examples/vehicle-scene-ledger.md](examples/vehicle-scene-ledger.md)\n- Capacity ledger template: [examples/capacity-ledger-template.md](examples/capacity-ledger-template.md)\n- Capacity conflict example: [examples/capacity-conflict-draft.zh-CN.md](examples/capacity-conflict-draft.zh-CN.md)\n- Draft audit example: [examples/problem-car-scene-draft.md](examples/problem-car-scene-draft.md)\n\n## Relationship Stance Continuity\n\nFor scenes with rival factions, secret relationships, hierarchy, family politics,\noffice politics, or sect leaders, use `--deep-review` or add `relationship-stance-audit`.\nIt extracts each dialogue line as `speaker -> listener/audience -> referenced party`\nand checks whether praise, criticism, comparison, naming, secrecy, and rank fit\nthe established relationship graph.\n\n- Guide: [docs/relationship-stance-continuity.md](docs/relationship-stance-continuity.md)\n- Ledger template: [examples/relationship-stance-ledger.zh-CN.md](examples/relationship-stance-ledger.zh-CN.md)\n\n## Character- and Situation-Fit Dialogue\n\n**Bidirectional interaction review addresses missing reactions and unfinished actions.**\nIt covers speech, wordless exchanges, and the viewpoint character's own responses.\nTrack initiator, affected recipient, separate speech/action obligations, evidence of\nreception, and completed or pending state. A spoken answer may leave a simultaneous\naction unfinished; one person's movement does not establish that another followed.\nRefusal, unawareness, stillness, or delayed uptake can be valid. Clearly implied\ncompletion needs no added gesture or emotional commentary.\n\nFiction and webnovel base Skills retain a brief check; detailed review uses the\nexisting on-demand `voice` profile. Automatic pipelines and deep chunk reviews also\nroute directed interpersonal actions. Use `--profile voice` when automatic cues miss\na language or phrasing. Reconciliation checks responses across chunk boundaries.\nThese are model-executed contextual checks; routing tests and deterministic lint do\nnot establish that a model will catch every omitted reaction.\n\n`dialogue-voice-audit` and `dialogue-performance-audit` separate stable speaker baseline,\nsituation-driven modulation, and the action each turn is trying to perform. Occupation, class, region, and trait\nlabels supply possible knowledge, incentives, duties, and register pressure; they do\nnot substitute for personality. An explicit speech or interpersonal-action task activates\nthe module on demand. Review an existing scene with an independent `voice` pass:\n\n```powershell\nhuman-writing-skills audit `\n  --draft my-dialogue-scene.md `\n  --context my-novel-ledger.md `\n  --profile voice\n```\n\nThe audit separates contradiction from motivated contrast and checks scene purpose,\nknowledge boundaries, practical constraints, response linkage, audience, and power.\nA consequential line or action does not require a mechanical spoken reply, but it\nmust receive verbal, physical, silently legible, interrupted, or deliberately deferred\nuptake before the prose shifts away. A physical beat is retained only when it changes\naccess, attention, leverage, permission, distance, or the next available action; the\nmodule does not prescribe touch, gestures, weather, clothing, or scenery after every line.\n\nIf the draft already exists, use `audit`:\n\n```powershell\npython -m humanwriting.cli audit `\n  --draft examples/problem-car-scene-draft.md `\n  --context examples/vehicle-scene-ledger.md\n```\n\n## Project Layout\n\n```text\nhumanwriting/        Python package and CLI\nskills/              reusable writing SKILLS in Markdown\nexamples/            sample continuity ledgers and article briefs\ntests/               standard-library unit tests\n```\n\n## CLI Usage\n\n### Optional Narrative Modules\n\nThe narrative controls use progressive disclosure. Generation adds\nthe dialogue modules only when a fiction or webnovel task explicitly asks for a\nspeech-centered or character-interaction scene such as dialogue, negotiation, reunion,\ntesting, reconciliation, confrontation, a meeting, interrogation, or argument.\nNarration-only and serious-document tasks do not trigger them. The `voice`,\n`serial`, `world`, `process`, `momentum`, `salience`, `recurrence`, `repetition`, `texture`, and\n`sources` and `preservation` audit profiles remain outside broad `full` review:\n\n```powershell\nhuman-writing-skills build --style fiction --task \"Write a negotiation in which both speakers want different outcomes.\"\nhuman-writing-skills build --style webnovel --context ledger.md --module serial-reentry --task \"Continue chapter 18.\"\nhuman-writing-skills audit --draft chapters.md --profile momentum\nhuman-writing-skills audit --draft chapter.md --profile texture\nhuman-writing-skills audit --draft chapter.md --profile process\nhuman-writing-skills audit --draft chapters.md --profile recurrence\n```\n\nThe dialogue modules model baseline speech, practical incentives, knowledge limits,\nscene goals, response linkage, and purposeful performance beats without treating a job as a personality. `dialogue-voice-audit` owns response obligation and deferred interaction debt; `dialogue-performance-audit` only tests whether a selected physical beat earns its place. Use\n`speech-register-continuity` for evidence-backed language identity, particles,\nhonorifics, and code-switching; use `capability-state-audit` for power and resource\nstate. Use `serial-reentry` only with\nprior chapters or a ledger, `momentum` for a multi-chapter draft, and `texture` for\nnarrative distance, cinematic opening stacks, imagery load, paragraph fragmentation,\nemotional over-explanation, and detail inventory. Use `world` only with explicit\nsetting constraints, `process` for consequential domain work, `salience` for long\ndrafts, `recurrence` for at least three chapters, `repetition` for long enough narrative\ndrafts with matching repetition/exposition cues, and `sources` only with serious\ndocuments and factual source files.\n\nDuring generation, world, process, and attention-budget modules activate only from\nexplicit setting, consequential-process, expansion, long-form, or dilution signals;\nordinary `--deep-review` does not load them.\nThe repetition/exposition module is likewise audit-first: use `--profile repetition`\nor narrative `ai-trace`, and let `pipeline --auto` add its separate pass only when\nthe draft supplies matching cues.\n\n### Audit Profiles\n\n`audit` can load only the checks needed for the current pass:\n\n| Profile | Purpose |\n| --- | --- |\n| `full` | Broad default audit; high-cost and strongly gated profiles remain separate |\n| `logic` | Cause, timeline, knowledge, motive, rules, resources, and consequences |\n| `character` | Character goal, voice, competence, boundaries, and change gates |\n| `voice` | Speaker baseline, scene goal, role/knowledge limits, audience register, change gates, and response obligations |\n| `register` | Language identity, dialect exposure, honorifics, particles, vocabulary, and code-switch gates |\n| `capability` | Power, skill, authority, equipment, injury, resources, counters, and transition gates; requires `--context` |\n| `serial` | Recap dumps, missing carryovers, and chapter resets; requires `--context` |\n| `momentum` | Multi-chapter entry pressure, irreversible turns, payoff, residue, and exit pressure |\n| `world` | Era, technology, institution, social-practice, and world-rule compatibility |\n| `process` | Promise, attempt, resistance, judgment, cost, evidence, and earned result |\n| `salience` | Long-draft attention allocation, dilution, and semantic echoes |\n| `recurrence` | Chapter fingerprints and repeated architecture across three or more chapters |\n| `repetition` | Narrative-only exact echoes, repeated action frames, and explanation that duplicates visible evidence |\n| `texture` | Narrative distance, scene-entry load, imagery, paragraph cadence, and detail disclosure |\n| `physical` | Position, capacity, reach, clothing, props, and injuries |\n| `relationship` | Audience, stance, information permissions, rank, and secret leaks |\n| `ai-trace` | Cliches, formulaic structure, static paragraphs, and other AI traces |\n| `ending` | Last meaningful change, reflective bookends, false closure, and genre-specific ending function |\n| `numbers` | False precision in action and emotion |\n| `proofread` | Omissions, sentence slots, stranded connectors, references, punctuation, naming, and layout |\n| `fidelity` | Meaning, entity, polarity, uncertainty, chronology, attribution, and invention checks; requires `--original` |\n| `preservation` | Useful ambiguity, repetition, motifs, hesitation, subtext, and speaker identity; requires `--original` and explicit selection |\n| `style-match` | Drift from explicitly supplied reference material; unavailable without a reference signal |\n| `sources` | Claim grounding against factual sources; requires a serious document and `--source` |\n\nProfiles can be combined, for example `--profile relationship --profile ai-trace`.\n\nOrdinary generation loads only a lightweight sentence-completeness guard. Full\nomission, missing-object, stranded-connector, and reference checks load only in the\n`proofread` profile or pipeline proofreading stage, preserving the generation token budget.\n\n### Chunked Long-Form Audit\n\nUse `chunk-audit` when a manuscript exceeds one reliable context window or was written\nacross model, prompt, or time changes. It complements `pipeline`: chunking handles\nmanuscript size and cross-block drift, while the pipeline separates different review\nresponsibilities for one draft.\n\n```powershell\nhuman-writing-skills chunk-audit `\n  --draft full-novel.md `\n  --style fiction `\n  --outline novel-outline.md `\n  --reference approved-sample.md `\n  --output-dir novel-consistency-audit\n```\n\nWithout an explicit reference, `--baseline-chunk` selects a candidate manuscript block;\napprove or correct it during baseline extraction. Reference prose supplies style evidence only. Fiction uses the\noutline or ledger for character canon and permits earned development; serious reports\nprotect facts, numbers, terminology, attribution, and conclusion scope. Default body,\ncontext, and baseline budgets keep the workflow usable on smaller-context models.\n\n### Verified Agent Review For Long Documents\n\nAdd `--agent-mode deep` when coverage matters more than token cost. It writes an explicit\ntask graph in `agent-plan.json`, requires a Coverage Receipt from every reviewer, and\nreserves report paths under `reports/`. After the baseline is approved, tasks that d\n\nArchive v0.15.1: 150 files, 328809 bytes\n\nFiles: agents/openai.yaml (435b), CONTRIBUTING.md (1300b), docs/agent-orchestration.md (5277b), docs/agent-orchestration.zh-CN.md (4819b), docs/audit-pipeline.md (6033b), docs/audit-pipeline.zh-CN.md (6504b), docs/chatbox.md (7460b), docs/chatbox.zh-CN.md (7173b), docs/editing-tools.md (4202b), docs/editing-tools.zh-CN.md (3869b), docs/forum-complaint-research.md (4820b), docs/forum-complaint-research.zh-CN.md (4977b), docs/ledger-extraction.md (1396b), docs/ledger-extraction.zh-CN.md (1290b), docs/long-form-consistency.md (5189b), docs/long-form-consistency.zh-CN.md (6247b), docs/number-sense.md (2079b), docs/number-sense.zh-CN.md (2336b), docs/pattern-linter.md (5135b), docs/pattern-linter.zh-CN.md (4942b), docs/physical-continuity.md (3075b), docs/physical-continuity.zh-CN.md (3753b), docs/protected-content.md (2004b), docs/protected-content.zh-CN.md (1984b), docs/reference-style.md (1859b), docs/reference-st...","readmeExcerpt":"Skill: Advanced Human Writing & AI Humanizer Owner: whh110112 Summary: AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity Tags: latest:0.15.3 Version history: v0.15.3 | 2026-09-29T12:02:20.699Z | user 中文 - 新增多语言改写保真审查，提示事实、归因、否定、范围等内容变化。 - 新增长篇作品按章节的整体结构审查，并完善覆盖回执。 - 增加多语言评测样例、文档及 CLI/MCP 测试。 - 保持按需加载，避免无关模块增加提示词负担。 English - Add multilingual rewrite-fidelity checks for chan","codeSnippets":[],"executableExamples":[{"language":"powershell","snippet":"human-writing-skills build --style fiction --context ledger.md --task \"Continue the scene.\"\nhuman-writing-skills humanize --draft chapter.md --style fiction --mode quick\nhuman-writing-skills humanize --draft article.md --style self-media --mode deep --with-examples\nhuman-writing-skills build --style fiction --reference sample.md --task \"Match the sample's restrained rhythm.\"\nhuman-writing-skills build --style self-media --original original.md --task \"Rewrite without adding facts.\"\nhuman-writing-skills audit --draft chapter.md --context ledger.md --profile physical\nhuman-writing-skills audit --draft chapter.md --profile voice\nhuman-writing-skills audit --draft chapter.md --context ledger.md --profile serial\nhuman-writing-skills audit --draft chapters.md --profile momentum\nhuman-writing-skills audit --draft chapters.md --profile recurrence\nhuman-writing-skills audit --draft chapter.md --profile process\nhuman-writing-skills audit --draft chapter.md --document-type fiction --profile ending\nhuman-writing-skills audit --draft paper.md --document-type academic-paper --source study.md --profile sources\nhuman-writing-skills audit --draft revised.md --original original.md --profile fidelity\nhuman-writing-skills audit --draft revised.md --original original.md --profile preservation\nhuman-writing-skills pipeline --draft chapter.md --context ledger.md --auto --with-stats --output-dir audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline outline.md --output-dir novel-audit\nhuman-writing-skills chunk-audit --draft full-novel.md --style fiction --outline outline.md --agent-mode deep --output-dir novel-agent-audit\nhuman-writing-skills verify-chunk-audit --package-dir novel-agent-audit\nhuman-writing-skills chunk-audit --draft report-es.md --style news-report --source sources.md --translationese --agent-mode deep --output-dir report-audit\nhuman-writing-mcp --root C:\\writing-project\nhuman-writing-skills lint --draft chapter.md --style fiction\nhuman-writ"},{"language":"powershell","snippet":"pip install human-writing-skills"},{"language":"powershell","snippet":"pip install -e ."},{"language":"powershell","snippet":"dsh plugin --profile web add dsh-advanced-human-writing"},{"language":"powershell","snippet":"human-writing-mcp --root C:\\writing-project"},{"language":"powershell","snippet":"human-writing-skills extract-ledger --draft chapters-01-10.md --context novel-ledger.md --output ledger-extraction-prompt.md"}],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"---\nname: human-writing-skills\ndescription: Advanced multilingual AI humanizer for natural rewriting, fiction editing, long-form audit and continuity, verified chunked agent review, translationese review, character consistency, opt-in whole-book story architecture review, and semantic rewrite-fidelity triage. Humanize AI text, remove robotic tone, edit fiction and novels, continue webnovel chapters, proofread writing, and audit story continuity, character voice, dialogue register and performance, scene geography, relationships, numbers, citations, source meaning, and translated-text fidelity. Use for AI writing cleanup, supplied-sample style matching, long-context fiction, essays, news, official, academic, legal, and technical prose. Trigger on humanize AI text, de-AI writing, natural rewriting, novel writing assistant, story consistency checker, scene ending audit, reflective ending, chunked audit, long-form agent audit, whole-book review, translationese audit, style consistency review, character consistency audit, dialogue audit, dialogue action audit, 增强版去 AI 写作 Skill、高级 AI 写作工具、去AI味、去AI写作、消除AI腔、AI人性化改写、AI文本润色、AI文章润色、小说润色、小说续写、AI式结尾、生硬结尾审查、无意义升华、长文一致性、长篇审查、分块审查、全书审查、文风统一、统一文风、人物设定统一、人物一致性审查、跨章一致性、小说审查、报告审查、人物口吻、人物对白审查、对话生硬、对白动作、方言语域、翻译腔、翻译审查、战力设定、场景空间审查、语义保真审查.\n---\n\n# Advanced Human Writing & AI Humanizer\n\nHumanize AI-shaped text, write or continue genre-aware prose, and audit long-form\ncontinuity without flattening a specific voice. Use the smallest set of modules\nthat covers the task. Keep project facts, prior chapters, rewrite originals, and\ncontinuity ledgers separate from optional style references.\n\n## Capability Layers\n\n- The `humanwriting/` Python package is executable. Its CLI deterministically\n  compiles prompts, locates recurring text patterns, calculates diagnostics,\n  previews conservative fixes, checks protected content, and writes staged audits.\n- The Markdown files under `skills/` are model-executed writing and editorial\n  modules. They are selected by the compiler and are intentionally not pretending\n  to be deterministic NLP algorithms.\n- A normal installation includes both layers. Verify the executable layer with\n  `human-writing-skills list --kind module` and run a real draft through `lint`,\n  `fix`, `verify`, or `pipeline` rather than judging the package from `SKILL.md` alone.\n- `human-writing-mcp` is an optional project-local coordination layer. It gives\n  separate agents bounded long-form assignments, stores receipts, and refuses\n  reconciliation until coverage verification passes. It does not call a model.\n\n## Quick Humanize Route\n\n- For a supplied draft, use `humanize --mode quick` for surface patterns,\n  rewrite fidelity, and voice/ambiguity preservation.\n- Use `--mode deep` only when structural repetition, cliches, paragraph stagnation,\n  or broad editorial reconstruction needs a separate pass.\n- Load `humanize-examples` only after an explicit request or `--with-examples`.\n- Do not activate rewrite-preservation module"},{"path":"plugins/deepseek-harness/README.md","content":"# Advanced Human Writing for DeepSeek Harness\n\nThis DeepSeek Harness bundle mounts the `human-writing-mcp` tools from the\nparent repository. It coordinates bounded long-form reviews locally: create a\nplan, let separate agents claim focused work, require complete coverage\nreceipts, and unlock the final reconciliation only after verification passes.\n\n## Prerequisites\n\nInstall the Python package in the environment selected by `PYTHON` or `python`:\n\n```powershell\npip install human-writing-skills\n```\n\nFor development from this repository:\n\n```powershell\npip install -e .\n```\n\n## Install\n\nAfter publishing this folder to npm, add it to a DeepSeek Harness profile:\n\n```powershell\ndsh plugin --profile web add dsh-advanced-human-writing\n```\n\nRestart the affected DSH profile. The bundle uses the workspace as its allowed\nfile root. Set `PYTHON` before launching DSH when the desired interpreter is\nnot named `python`.\n\nThe mounted tools are `plan_long_form_audit`, `list_audit_tasks`,\n`claim_audit_task`, `submit_audit_report`, `verify_audit_coverage`,\n`get_reconciliation_task`, and `read_project_context`.\n\nThis package does not send drafts to a third party and does not call a model by\nitself. The active DSH agent performs each assigned review and must submit its\nown report."},{"path":"README.md","content":"# Advanced Human Writing & AI Humanizer\n\n> Reusable multilingual writing `SKILLS` for natural prose, genre-aware style, and long-form continuity.\n\n**Advanced AI humanizer and de-AI writing toolkit** for natural rewriting,\nAI text cleanup, fiction editing, novel continuation, chunked long-form audit,\nwriting style unification, and character consistency review.\n\n[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)\n[![Python](https://img.shields.io/badge/Python-3.9%2B-blue.svg)](pyproject.toml)\n[![Zero Dependencies](https://img.shields.io/badge/dependencies-zero-brightgreen.svg)](pyproject.toml)\n\n[中文说明](README.zh-CN.md) | English | [Español](README.es.md) | [Português (Brasil)](README.pt-BR.md) | [Français](README.fr.md)\n\nAdvanced Human Writing & AI Humanizer is an open-source, modular skill pack and lightweight prompt compiler for natural multilingual AI-assisted writing. The package, repository, and ClawHub slug remain `human-writing-skills` for compatibility.\n\nThis is not an empty prompt collection. The repository contains two deliberately\ndifferent capability layers:\n\n- **Executable Python tools:** `humanwriting/` provides the installable\n  `human-writing-skills` CLI for deterministic lint findings with evidence spans,\n  text statistics, conservative fix previews, protected-content verification,\n  prompt compilation, and staged audit-file generation.\n- **Model-executed editorial modules:** `skills/*.md` contains genre and review\n  instructions selected on demand by that compiler. These modules guide the writing\n  model; they do not falsely present subjective literary judgment as a deterministic\n  algorithm.\n\nThe test suite exercises both the executable layer and module-selection gates.\n\n## Agent Orchestration, MCP, And DeepSeek Harness Plugin\n\nFor book-length novels, report series, or research coverage that cannot be\ntrusted to one chat window, install the dependency-free `human-writing-mcp`\nserver. It makes the long-form task graph executable across Codex, Claude Code,\nOpenCode, DeepSeek Harness, Manus, and Hermes: agents claim bounded tasks,\nsubmit coverage receipts, and cannot unlock reconciliation until every required\nreview is complete. The service stays inside a chosen project root and does not\ncall a model or upload drafts.\n\n```powershell\nhuman-writing-mcp --root C:\\writing-project\n```\n\nThe repository also ships a native DeepSeek Harness npm/Cordis bundle under\n[`plugins/deepseek-harness`](plugins/deepseek-harness). It mounts the official\nDSH MCP client and starts the same verified local coordination service. See the\n[agent orchestration guide](docs/agent-orchestration.md) and\n[DeepSeek Harness plugin guide](plugins/deepseek-harness/README.md).\n\nThe MCP server also supports day-to-day single-document work: `lint_text`,\n`get_style_statistics`, `verify_protected_content`, `verify_fidelity`, `compile_humanize_prompt`,\n`compile_audit_prompt`, and `compile_ledger_extraction`. It advertises a small\nnative prompt menu ("},{"path":"_meta.json","content":"{\n  \"ownerId\": \"kn7cv6zdh5xrvfc30xqepze49582aw5n\",\n  \"slug\": \"human-writing-skills\",\n  \"version\": \"0.15.3\",\n  \"publishedAt\": 1790683340699\n}"},{"path":"CONTRIBUTING.md","content":"# Contributing\n\nThank you for improving Advanced Human Writing & AI Humanizer.\n\n## Rules And Tests\n\n- Add a deterministic rule only when it has a concrete, user-visible editing purpose.\n- Give every rule a stable ID, category, severity, repair direction, allow-list behavior,\n  positive fixture, and a plausible negative fixture.\n- Do not describe a pattern score as proof of AI authorship. Rules identify editing leads.\n- Gate genre-specific rules so fiction, news, academic, legal, and technical writing do\n  not inherit one another's constraints.\n- Preserve token budgets: optional or specialized modules must not silently enter quick\n  generation or unrelated document types.\n\nRun the full suite before opening a pull request:\n\n```powershell\npython -m unittest discover -s tests -v\npython -m build\n```\n\n## Documentation\n\nUpdate the English and Chinese README material for user-facing commands. Keep examples\ngeneric and evidence-led; do not publish private manuscript material in fixtures or issues.\n\n## Pre-commit\n\nTeams can install the repository hook with:\n\n```powershell\npre-commit install --hook-type pre-commit\n```\n\nThe hook checks Markdown-like text and fails only once the transparent pattern score\nreaches its configured threshold. Use allow-lists for intentional motifs or house style."}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":"AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity Skill: Advanced Human Writing & AI Humanizer Owner: whh110112 Summary: AI text humanizer for de-AI writing, natural rewriting, fiction editing, and novel continuity Tags: latest:0.15.3 Version history: v0.15.3 | 2026-09-29T12:02:20.699Z | user 中文 - 新增多语言改写保真审查，提示事实、归因、否定、范围等内容变化。 - 新增长篇作品按章节的整体结构审查，并完善覆盖回执。 - 增加多语言评测样例、文档及 CLI/MCP 测试。 - 保持按需加载，避免无关模块增加提示词负担。 English - Add multilingual rewrite-fidelity checks for chan","editorialQuality":{"score":100,"threshold":65,"status":"ready","wordCount":1914,"uniquenessScore":47,"reasons":[]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":"No screenshots, media assets, or demo links are available."},"primaryImageUrl":null,"mediaAssetCount":0,"assets":[],"demoUrl":null},"ownerResources":{"evidence":{"source":"unclaimed","verified":false,"confidence":"low","updatedAt":"2026-10-10T02:23:34.856Z","emptyReason":"This page has not been claimed by the agent owner."},"hasCustomPage":false,"customPageUpdatedAt":null,"customLinks":[],"structuredLinks":{"docsUrl":null,"demoUrl":null,"supportUrl":null,"pricingUrl":null,"statusUrl":null},"customPage":null},"relatedAgents":{"evidence":{"source":"protocol-neighbors","verified":false,"confidence":"medium","updatedAt":"2026-10-10T08:48:30.026Z","emptyReason":null},"items":[{"id":"8ebccd8e-3863-4187-8355-c3f14e1f9edf","entityType":"agent","canonicalPath":"/agent/iofficeai-aionui","slug":"iofficeai-aionui","name":"AionUi","description":"Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!","url":"https://github.com/iOfficeAI/AionUi","homepage":"https://www.aionui.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-10-09T19:11:12.944Z","createdAt":"2026-02-25T03:38:16.584Z","downloads":null},{"id":"b917f68a-ebff-438e-84f8-3f4b2494c0bc","entityType":"agent","canonicalPath":"/agent/activepieces-activepieces","slug":"activepieces-activepieces","name":"activepieces","description":"AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents","url":"https://github.com/activepieces/activepieces","homepage":"https://www.activepieces.com","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-15T02:22:12.426Z","createdAt":"2026-02-25T03:38:12.412Z","downloads":null},{"id":"5cb26759-3a39-483f-94cf-276a98c13bb8","entityType":"agent","canonicalPath":"/agent/cherryhq-cherry-studio","slug":"cherryhq-cherry-studio","name":"cherry-studio","description":"AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs","url":"https://github.com/CherryHQ/cherry-studio","homepage":"https://cherry-ai.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-11T14:38:40.986Z","createdAt":"2026-02-25T03:38:19.379Z","downloads":null},{"id":"6f6582d0-5d76-4f0f-b81d-86520247950b","entityType":"agent","canonicalPath":"/agent/copilotkit-copilotkit","slug":"copilotkit-copilotkit","name":"CopilotKit","description":"The Frontend for Agents & Generative UI. React + Angular","url":"https://github.com/CopilotKit/CopilotKit","homepage":"https://docs.copilotkit.ai","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-03-25T09:50:57.846Z","createdAt":"2026-02-25T03:39:14.617Z","downloads":null}],"links":{"hub":"/agent","source":"/agent/source/clawhub","protocols":[{"label":"OpenClaw","href":"/agent/protocol/openclew"}]}}}