{"id":"f06ae4e0-b0ee-40d1-b148-f8eb0b03b64d","entityType":"agent","slug":"clawhub-c-narcissus-paper-framework-figure-studio-pro","name":"Paper Framework Figure Studio Pro","canonicalUrl":"https://www.xpersona.co/agent/clawhub-c-narcissus-paper-framework-figure-studio-pro","canonicalPath":"/agent/clawhub-c-narcissus-paper-framework-figure-studio-pro","generatedAt":"2026-10-10T17:35:53.061Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-10T15:27:10.104Z","emptyReason":null},"description":"Use when the user wants to design, prompt, generate, critique, or integrate publication-ready research-paper framework figures: method overview diagrams, arc...","descriptionLabel":"Source description","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 1.4K downloads reported by the source. Last updated 10/10/2026.","installCommand":"clawhub skill install s177jw27anm32qt3dm1c7dq0jh851mjz:paper-framework-figure-studio-pro","sourceUrl":"https://clawhub.ai/c-narcissus/paper-framework-figure-studio-pro","homepage":"https://clawhub.ai/c-narcissus/skills/paper-framework-figure-studio-pro","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/c-narcissus/paper-framework-figure-studio-pro","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/c-narcissus/skills/paper-framework-figure-studio-pro","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":63,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Paper Framework Figure Studio Pro technical dossier on Xpersona with agent coverage, OPENCLEW support, and live trust metadata."},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-10T15:27:10.104Z","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":"2026-10-10T15:27:10.104Z","emptyReason":null},"stars":null,"forks":null,"downloads":1361,"packageName":null,"latestVersion":"1.2.0","tractionLabel":"1.4K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-10T15:27:10.089Z","emptyReason":null},"lastUpdatedAt":"2026-10-10T15:27:10.104Z","lastCrawledAt":"2026-10-10T15:27:10.089Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-11T15:27:10.089Z","lastVerifiedAt":null,"highlights":[{"version":"1.2.0","createdAt":"2026-05-05T12:58:04.848Z","changelog":"**Version 1.2.0 – Major workflow and structure upgrade** - Overhauled workflow for strict step-by-step, multi-turn process: now separates startup, text candidates, visual candidate setup, image generation, and review into distinct reply types. - Enforces a mandatory candidate-image bridge: after text options, a visual candidate board (normally 6 images) is required before final image selection. - Implements strict separation of text and image replies; no mixing of planning text and image generation in the same reply. - Adds detailed state tracking and required state footers in every text reply, showing all steps, current position, mode, outputs, and pending actions. - Startup now outputs only a startup plan, never immediately generates or analyzes—users must confirm or supply materials first. - Expanded examples, references, and templates via significant file additions and removals to support new workflow and evidence corpus.","fileCount":33,"zipByteSize":40768},{"version":"1.0.3","createdAt":"2026-04-23T12:05:05.161Z","changelog":"- Added license information to SKILL.md (MIT-0). - Added OpenClaw-specific metadata under `openclaw` in SKILL.md. - Moved compatibility information into metadata. - Added missing agents/openai.yaml file.","fileCount":22,"zipByteSize":38935},{"version":"1.0.2","createdAt":"2026-04-23T11:39:02.489Z","changelog":"No functional or content changes in this version. - Version bumped from 1.4.2 to 1.4.3 in skill metadata. - No other modifications detected.","fileCount":21,"zipByteSize":38480},{"version":"1.0.1","createdAt":"2026-04-23T11:24:47.850Z","changelog":"**Paper Framework Figure Studio Pro v1.4.2** Major update: Improved workflow structure, explicit state management, and detailed setup/usage guidance for multi-round figure design. - Added robust documentation files (templates, examples, references, publishing guides) to support consistent multi-round, stateful workflow. - Enforced strict turn-separation protocol: all planning and approval steps are text-only; image rendering occurs only after explicit user confirmation. - Introduced first-contact and first-turn protocols to recommend but not require upstream deep-reading reports, guiding users on preferred input and next steps. - Clarified non-negotiable rules around human approval, state tracking, and workflow for image candidate generation and selection. - Updated compatibility metadata and improved clarity on workflow expectations, next step guidance, and decision points. - Removed legacy and redundant documentation to streamline onboarding and publishing.","fileCount":21,"zipByteSize":36729},{"version":"1.0.0","createdAt":"2026-04-23T05:17:09.055Z","changelog":"Paper Framework Figure Studio Pro is a multi-round scientific diagram co-designer specialized for paper framework diagrams and method overview figures in top-tier computer science papers. It helps transform a paper deep-reading report, method description, or model overview into publication-ready figures through an iterative human-in-the-loop workflow. Instead of producing a single figure in one shot, the skill guides the user through structured rounds of framework-figure design: first selecting a high-level figure family, then narrowing down substyles, layout, information density, and local visual elements. After each round, it generates multiple candidate diagrams, lets the user choose or combine directions, records preferences, and updates the design state before continuing. The skill is optimized for reviewer communication efficiency. It prioritizes clear problem framing, framework structure, module interactions, key novelty, and why-the-method-works comparisons. It supports several figure families commonly seen in recent top-tier CS papers, including classic framework overview, modern modular card layouts, mechanism-explanation figures with intermediate snapshots, and design-forward scientific illustration styles. After generating the final figure, the skill explicitly asks whether the user also wants supporting figure text, such as short in-figure labels, a polished figure caption, panel-by-panel explanation text, or slide-friendly annotation text.","fileCount":8,"zipByteSize":7127}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s177jw27anm32qt3dm1c7dq0jh851mjz:paper-framework-figure-studio-pro","setupComplexity":"low","setupSteps":["Install using `clawhub skill install s177jw27anm32qt3dm1c7dq0jh851mjz:paper-framework-figure-studio-pro` in an isolated environment before connecting it to live workloads.","No published capability contract is available yet, so validate auth and request/response behavior manually.","Review the upstream CLAWHUB listing at https://clawhub.ai/c-narcissus/paper-framework-figure-studio-pro before using production credentials."],"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-c-narcissus-paper-framework-figure-studio-pro/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-10T17:35:53.057Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-c-narcissus-paper-framework-figure-studio-pro/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":"medium","updatedAt":"2026-10-10T15:27:10.104Z","emptyReason":null},"readme":"Skill: Paper Framework Figure Studio Pro\n\nOwner: c-narcissus\n\nSummary: Use when the user wants to design, prompt, generate, critique, or integrate publication-ready research-paper framework figures: method overview diagrams, arc...\n\nTags: latest:1.2.0\n\nVersion history:\n\nv1.2.0 | 2026-05-05T12:58:04.848Z | user\n\n**Version 1.2.0 – Major workflow and structure upgrade**\n\n- Overhauled workflow for strict step-by-step, multi-turn process: now separates startup, text candidates, visual candidate setup, image generation, and review into distinct reply types.\n- Enforces a mandatory candidate-image bridge: after text options, a visual candidate board (normally 6 images) is required before final image selection.\n- Implements strict separation of text and image replies; no mixing of planning text and image generation in the same reply.\n- Adds detailed state tracking and required state footers in every text reply, showing all steps, current position, mode, outputs, and pending actions.\n- Startup now outputs only a startup plan, never immediately generates or analyzes—users must confirm or supply materials first.\n- Expanded examples, references, and templates via significant file additions and removals to support new workflow and evidence corpus.\n\nv1.0.3 | 2026-04-23T12:05:05.161Z | user\n\n- Added license information to SKILL.md (MIT-0).\n- Added OpenClaw-specific metadata under `openclaw` in SKILL.md.\n- Moved compatibility information into metadata.\n- Added missing agents/openai.yaml file.\n\nv1.0.2 | 2026-04-23T11:39:02.489Z | user\n\nNo functional or content changes in this version.\n\n- Version bumped from 1.4.2 to 1.4.3 in skill metadata.\n- No other modifications detected.\n\nv1.0.1 | 2026-04-23T11:24:47.850Z | user\n\n**Paper Framework Figure Studio Pro v1.4.2**\n\nMajor update: Improved workflow structure, explicit state management, and detailed setup/usage guidance for multi-round figure design.\n\n- Added robust documentation files (templates, examples, references, publishing guides) to support consistent multi-round, stateful workflow.\n- Enforced strict turn-separation protocol: all planning and approval steps are text-only; image rendering occurs only after explicit user confirmation.\n- Introduced first-contact and first-turn protocols to recommend but not require upstream deep-reading reports, guiding users on preferred input and next steps.\n- Clarified non-negotiable rules around human approval, state tracking, and workflow for image candidate generation and selection.\n- Updated compatibility metadata and improved clarity on workflow expectations, next step guidance, and decision points.\n- Removed legacy and redundant documentation to streamline onboarding and publishing.\n\nv1.0.0 | 2026-04-23T05:17:09.055Z | user\n\nPaper Framework Figure Studio Pro is a multi-round scientific diagram co-designer specialized for paper framework diagrams and method overview figures in top-tier computer science papers. It helps transform a paper deep-reading report, method description, or model overview into publication-ready figures through an iterative human-in-the-loop workflow.\n\nInstead of producing a single figure in one shot, the skill guides the user through structured rounds of framework-figure design: first selecting a high-level figure family, then narrowing down substyles, layout, information density, and local visual elements. After each round, it generates multiple candidate diagrams, lets the user choose or combine directions, records preferences, and updates the design state before continuing.\n\nThe skill is optimized for reviewer communication efficiency. It prioritizes clear problem framing, framework structure, module interactions, key novelty, and why-the-method-works comparisons. It supports several figure families commonly seen in recent top-tier CS papers, including classic framework overview, modern modular card layouts, mechanism-explanation figures with intermediate snapshots, and design-forward scientific illustration styles.\n\nAfter generating the final figure, the skill explicitly asks whether the user also wants supporting figure text, such as short in-figure labels, a polished figure caption, panel-by-panel explanation text, or slide-friendly annotation text.\n\nArchive index:\n\nArchive v1.2.0: 33 files, 40768 bytes\n\nFiles: agents/openai.yaml (2586b), CHANGELOG.md (1020b), examples/example-candidate-review-turn.md (1539b), examples/example-final-caption-package.md (1641b), examples/example-image-only-candidate-board-turn.md (520b), examples/example-routing-turn.md (1762b), examples/example-startup-turn.md (2357b), examples/example-text-candidates-turn.md (1360b), examples/example-visual-board-turn.md (1785b), metadata.json (6721b), publish/listing_long.md (2425b), publish/listing_short.md (645b), publish/release_checklist.md (1792b), publish/starter_messages.md (1217b), README.md (5207b), references/builder-time-acquisition-report.md (1180b), references/evidence-lineage-summary.md (2331b), references/evidence-map-index.md (1863b), references/figure-class-taxonomy.md (2839b), references/figure-pattern-library.md (2125b), references/initial-corpus-manifest.md (1115b), references/prompt-generation-policy.md (2235b), references/review-rubric.md (1849b), references/source-corpus-notes.md (2091b), references/visual-style-and-board-protocol.md (1601b), references/workflow-and-state-contract.md (2641b), skill-card.md (3319b), SKILL.md (10784b), templates/figure-brief-template.md (1132b), templates/prompt-template.md (1244b), templates/state-footer-template.md (2196b), templates/user-input-bundle.md (1177b), _meta.json (152b)\n\nFile v1.2.0:SKILL.md\n\n---\nname: paper-framework-figure-studio-pro\nlicense: MIT-0\ndescription: \"Use when the user wants to design, prompt, generate, critique, or integrate publication-ready research-paper framework figures: method overview diagrams, architecture diagrams, pipeline/process diagrams, agent workflows, system/data-flow figures, mechanism-intuition figures, case walkthrough panels, and reviewer-facing schematic figures. Generated from research-paper-figure-skill-factory v1.0.1 with full-feasible local PDF evidence, startup-plan-only first replies, strict text/image separation, mandatory text-candidate to visual-candidate setup to image-only candidate board to candidate-review selection workflow, optional sample images, ChatGPT web Create image / ChatGPT Images 2.0, Codex $imagegen first, all-step/current-position state footers, and next-question help in every text reply.\"\nmetadata:\n  display_name: Paper Framework Figure Studio Pro\n  version: \"1.2.0\"\n  author: OpenAI\n  tags:\n    - research-figure\n    - paper-framework\n    - method-framework\n    - architecture-diagram\n    - pipeline-diagram\n    - agent-workflow\n    - candidate-image-bridge\n    - imagegen\n    - chatgpt-images-2\n    - clawhub\n    - openclaw\n  compatibility: Codex, ChatGPT web, OpenClaw, ClawHub marketplace. Requires image-generation capability for rendering.\n  openclaw:\n    skillKey: paper-framework-figure-studio-pro\n---\n\n# Paper Framework Figure Studio Pro\n\nThis skill designs publication-ready raster framework figures for computer-science research papers. Use it for method overviews, architecture diagrams, pipelines, agent workflows, system/data-flow figures, mechanism-intuition figures, case walkthroughs, and reviewer-facing schematic figures.\n\nIt was regenerated with `research-paper-figure-skill-factory` v1.0.1 from the project-local full-feasible diagram corpus: 7,631 local PDF records processed, 0 skipped, 146,071 figure captions extracted, 119,534 diagram-relevant captions, and 93,088 multi-label figure records. Framework-relevant evidence includes method-framework, architecture, pipeline/process, agent-workflow, mechanism, and case-walkthrough patterns. Representative rendered pages are audit aids only, not the corpus size.\n\n## Non-Negotiable Contract\n\n### First Trigger\n\nOn the first reply in a new project, output only a startup plan. Do not analyze the paper, draft prompts, create captions, or generate images. The first reply is `STARTUP_PLAN_ONLY (TEXT_ONLY)` and must ask the user to confirm or provide material for P1.\n\nIf the first user message asks to \"直接出图\", \"生成 6 张图\", \"出候选图\", \"generate images\", or otherwise asks for image generation, record the request as pending only. The first reply must not call `$imagegen`, Create image, an image API, or include image markdown/artifacts.\n\n### Mandatory Candidate-Image Bridge\n\nAfter any multi-option text decision, do not move directly to final prompt, final image generation, caption, or text-only locking. Use this mandatory bridge:\n\n1. `TEXT_ONLY` text-candidate turn: present 4-6 text candidates, normally 6.\n2. `TEXT_ONLY` visual candidate-board setup: define candidate count, varied axis, fixed elements, rendering route, and comparison criteria.\n3. `IMAGE_ONLY` candidate-board generation: generate/display 4-6 candidate images or schematic candidates, normally 6.\n4. `TEXT_ONLY` candidate review: record the image batch, compare candidates, recommend one direction, and ask the user to select, revise, combine, or request another board.\n\nThis bridge is mandatory after candidate schemes, subtype choices, layout choices, style choices, metaphor choices, density choices, and prompt alternatives. Skip it only if the user explicitly says to stay text-only or skip image candidates, and then record `visual_candidate_board_skipped_by_user: true`.\n\n### Strict Text/Image Separation\n\nEvery assistant response is exactly one mode:\n\n- `TEXT_ONLY`: planning, intake, diagnosis, candidate text, candidate-board setup, prompt writing, critique, state update, and confirmation request.\n- `IMAGE_ONLY`: image generation only. No prose, caption, prompt text, critique, or state footer.\n\nIf a reply emits visible text, do not generate images in the same response. If generation is ready, ask for confirmation and stop. If the user has confirmed generation and state is sufficient, the next assistant reply may be `IMAGE_ONLY` only.\n\n### Rendering Route\n\nFor candidate boards, drafts, final diagrams, and revisions:\n\n1. In ChatGPT web, use **Create image** through **ChatGPT Images 2.0**.\n2. In Codex, use the `$imagegen` skill first.\n3. If `$imagegen` is unavailable in Codex, use ChatGPT Images 2.0 API or another approved image-generation API.\n4. Native bitmap outputs such as PNG, JPG, JPEG, or WebP are allowed.\n5. Do not use SVG, Mermaid, TikZ, Graphviz, HTML/CSS, canvas, matplotlib, filesystem code drawing, or code-rendered/exported images as candidate, draft, final, or fallback visuals.\n\n### Every Text Reply\n\nEvery `TEXT_ONLY` reply must include these sections in order:\n\n1. `当前执行计划`\n2. The substantive work for the current step\n3. `默认推荐`\n4. `当前状态与产物`\n5. `下一步你可以这样问`\n\nThe state footer must include `全部步骤与当前位置`, current response mode, current-turn outputs, cumulative outputs, pending outputs, candidate-board state, and the previous `IMAGE_ONLY` batch recording status.\n\nThe first copyable prompt must begin:\n\n`请使用**paper-framework-figure-studio-pro**，执行，根据当前状态，下一步执行：...`\n\nAlways include this fallback prompt:\n\n`请使用**paper-framework-figure-studio-pro**，根据当前状态，提供下一步提问建议。`\n\nNormal follow-up turns continue from the active session/history. Ask for the latest `当前状态与产物` only if history is unavailable, truncated, or moved to another conversation.\n\n## Required Workflow\n\n| Step | Reply Type | Goal | Output |\n|---|---|---|---|\n| S0 | STARTUP_PLAN_ONLY (TEXT_ONLY) | Startup confirmation only | Startup plan |\n| P1 | TEXT_ONLY | Intake target-paper material, target slot, constraints, and optional sample images | Material status |\n| P2 | TEXT_ONLY | Diagnose framework-figure need and multi-label subtype routing | Subtype candidates + default route |\n| P3 | TEXT_ONLY | Define reader effect and produce 4-6 text candidate schemes, normally 6 | Text candidates + required visual-candidate next action |\n| P4 | TEXT_ONLY | Set up visual candidate board: count, varied axis, fixed content, route, and comparison criteria | Candidate-board brief |\n| P5 | IMAGE_ONLY | Generate/display 4-6 candidate images or schematic candidates, normally 6 | Candidate images only |\n| P6 | TEXT_ONLY | Record the candidate image batch, compare candidates, recommend one, and lock or revise direction | Selected/revised visual direction |\n| P7 | TEXT_ONLY | Build final content architecture and formal image brief/prompt for the selected direction | Final image brief |\n| P8 | IMAGE_ONLY | Generate formal figure candidate or revision batch through the approved image route | Formal image candidates only |\n| P9 | TEXT_ONLY | Review, refine, caption, legend, body insertion, and handoff text | Final paper text package |\n\nP4/P5/P6 are not optional after P3 when multiple text options were presented. They are the visual selection bridge.\n\n## Candidate Defaults\n\n- Text candidates: 4-6, normally 6.\n- Candidate-board images: 4-6, normally 6.\n- Formal image candidates: 4-6, normally 6 unless a selected direction needs fewer variants.\n- If the user says only \"继续\", \"出图\", \"生成\", or \"generate\" after a text-candidate or board-setup turn, default to 6 candidate images.\n- Generate one image only when the user explicitly asks for one.\n- If a text reply presents multiple schemes, layouts, styles, metaphors, densities, or prompt options, the first recommended next prompt must ask to generate/display multiple candidate images or schematic candidates, normally 6.\n\n## Diagram Routing\n\nRecord all applicable labels before locking a primary production subtype. A single paper or diagram may belong to multiple classes.\n\nFramework-focused labels:\n\n- `method_framework`\n- `architecture`\n- `pipeline_process`\n- `agent_workflow`\n- `system_data_flow`\n- `mechanism_intuition`\n- `case_walkthrough`\n- `graph_network`\n- `evidence_board`\n- `taxonomy_design_space`\n- `data_benchmark_protocol`\n- `failure_limitation`\n- `theory_proof_intuition`\n- `general_diagram_or_figure`\n\nChoose one primary production subtype for the current rendering, but keep secondary labels as constraints on layout, arrows, labels, and density.\n\n## Sample / Reference Images\n\nSample images are optional. Ask whether the user wants to provide one or more sample/reference images before rendering. For each image, record the preferred transfer attributes:\n\n- style\n- layout\n- panel rhythm\n- information density\n- content-detail level\n- label style and label placement\n- color semantics\n- callout grammar\n- negative reference constraints\n\nDo not copy sample-image content, claims, data, identities, or proprietary marks unless the user explicitly owns or authorizes that content. Use samples as controllable visual references only.\n\n## State Fields\n\nPreserve these fields in every text reply:\n\n- current mode and current step\n- all workflow steps and current position\n- material status\n- paper thesis / figure thesis\n- diagram labels and primary production subtype\n- reader-effect contract\n- required modules, labels, and constraints\n- sample/reference image transfer map\n- text candidate count and candidate IDs\n- visual candidate-board status\n- visual board type, varied axis, fixed elements, candidate count\n- candidate image batch ID\n- visual candidate history and selected visual candidate\n- final image brief status\n- rendering route\n- current-turn outputs, cumulative outputs, pending outputs\n- whether the previous `IMAGE_ONLY` output has been recorded\n- next recommended action\n\nIf history is incomplete, do not invent missing state. Ask the user to provide the latest `当前状态与产物` or the missing material.\n\n## References\n\nUse these package references as needed:\n\n- `references/workflow-and-state-contract.md`\n- `references/visual-style-and-board-protocol.md`\n- `references/prompt-generation-policy.md`\n- `references/figure-class-taxonomy.md`\n- `references/figure-pattern-library.md`\n- `references/review-rubric.md`\n- `references/source-corpus-notes.md`\n- `references/evidence-map-index.md`\n- `references/evidence-lineage-summary.md`\n- `references/builder-time-acquisition-report.md`\n- `references/initial-corpus-manifest.md`\n- `templates/state-footer-template.md`\n- `templates/figure-brief-template.md`\n- `templates/prompt-template.md`\n- `templates/user-input-bundle.md`\n\nFile v1.2.0:README.md\n\n# Paper Framework Figure Studio Pro\n\n## 中文\n\n`paper-framework-figure-studio-pro` 用于帮助研究者为论文生成 framework diagram、method diagram、pipeline diagram、architecture diagram 和 agent workflow 等框架图。它适合把论文 PDF、摘要、方法说明或草稿转化为可比较的制图方案、候选图、修改建议、caption 和图注说明。\n\n### 推荐使用方式\n\n优先在 ChatGPT 网页版中使用，并选择 **Extended thinking**。网页版更适合完成完整的论文理解、候选图生成和多轮修图流程。\n\n如果接下来的步骤是生成图片，最好在 ChatGPT 网页版中手动选择 **Create image** 模式，再让它继续生成候选图或最终图。\n\n在 Codex 里也可以尝试使用，但可能会遇到图像生成或上下文处理问题，而且会比较费 token。除非你明确需要在本地工程目录中整理文件、改 skill 或生成配套文档，否则不建议把主要制图流程放在 Codex 里完成。\n\n### ChatGPT 网页版使用步骤\n\n1. 把 `paper-framework-figure-studio-pro-v1.2.0-skill.zip` 放进 ChatGPT 的 Sources。\n2. 把论文 PDF 也放进 Sources，例如 `semiDFL.pdf`。\n3. 选择 Extended thinking。\n4. 输入类似下面的 prompt：\n\n```text\n请严格按照 paper-framework-figure-studio-pro-v1.2.0-skill.zip 里 skill 的步骤，对 semiDFL.pdf 绘制 diagram。不要参考 semiDFL.pdf 里面已有的 diagram。\n```\n\n如果你的论文文件名不是 `semiDFL.pdf`，请把 prompt 里的文件名替换为实际上传到 Sources 的文件名。\n\n当 skill 已经完成文字方案比较，并提示下一步要生成候选图或最终图时，建议手动切换到 **Create image** 模式后再继续。\n\n### 制图流程\n\n1. 提供论文 PDF、摘要、方法说明、目标章节或已有草稿。\n2. 说明是否要避开论文中已有 diagram，以及是否提供参考图。\n3. Skill 先判断这张图更适合 framework、architecture、pipeline、workflow 还是 mechanism diagram。\n4. 先生成 4-6 个文字方案，通常是 6 个。\n5. 选择或确认候选图方向；如果有参考图，可以说明每张图只参考布局、风格、信息密度、标签或配色中的哪些属性。\n6. 生成多张候选图或示意图供比较。\n7. 从候选图中选择最接近的一张，或指出需要修改的地方。\n8. 根据选择继续生成正式版本或修订版本。\n9. 最后整理 caption、legend 和正文中的图说明文字。\n\n## English\n\n`paper-framework-figure-studio-pro` helps researchers create framework diagrams, method diagrams, pipeline diagrams, architecture diagrams, and agent workflows for research papers. It turns a paper PDF, abstract, method description, or draft notes into comparable diagram directions, candidate figures, revision guidance, captions, and figure descriptions.\n\n### Recommended Use\n\nPrefer using this skill in the ChatGPT web app with **Extended thinking** enabled. The web app is better suited for the full workflow: paper understanding, candidate figure generation, and iterative figure revision.\n\nIf the next step is image generation, it is best to manually select **Create image** mode in the ChatGPT web app before asking it to generate candidate figures or the final figure.\n\nYou can also try it in Codex, but image generation and context handling may be less reliable, and it can consume many tokens. Unless you specifically need local file organization, skill editing, or repository documentation, the main figure-making workflow is better done in ChatGPT web.\n\n### ChatGPT Web Usage\n\n1. Add `paper-framework-figure-studio-pro-v1.2.0-skill.zip` to ChatGPT Sources.\n2. Add the paper PDF to Sources as well, for example `semiDFL.pdf`.\n3. Select Extended thinking.\n4. Type a prompt like this:\n\n```text\nPlease strictly follow the workflow in paper-framework-figure-studio-pro-v1.2.0-skill.zip to draw a diagram for semiDFL.pdf. Do not refer to any existing diagram inside semiDFL.pdf.\n```\n\nIf your paper file is not named `semiDFL.pdf`, replace the file name in the prompt with the exact file name uploaded to Sources.\n\nWhen the skill has finished comparing text directions and the next step is to generate candidate figures or a final figure, manually switch to **Create image** mode before continuing.\n\n### Figure-Making Workflow\n\n1. Provide the paper PDF, abstract, method description, target section, or draft notes.\n2. Specify whether existing diagrams in the paper should be ignored and whether reference images are provided.\n3. The skill diagnoses whether the figure should be a framework, architecture, pipeline, workflow, or mechanism diagram.\n4. It first proposes 4-6 text directions, usually 6.\n5. You confirm the candidate-image direction. If reference images are provided, specify which attributes to borrow from each image, such as layout, style, information density, labels, or color.\n6. It generates multiple candidate figures or schematic candidates for comparison.\n7. You select the closest candidate or describe what needs to change.\n8. It then generates a formal version or revision based on the selected direction.\n9. It can finally draft the caption, legend, and in-paper figure description.\n\nFile v1.2.0:_meta.json\n\n{\n  \"ownerId\": \"kn7fxns1xpr6z67w885my7d7k98506vv\",\n  \"slug\": \"paper-framework-figure-studio-pro\",\n  \"version\": \"1.2.0\",\n  \"publishedAt\": 1777985884848\n}\n\nFile v1.2.0:references/builder-time-acquisition-report.md\n\n# Builder-Time Acquisition Report\n\nGenerated by `research-paper-figure-skill-factory` v1.0.1 for `paper-framework-figure-studio-pro` v1.2.0.\n\n## Source\n\nThe skill uses the previously extracted project-local full-feasible diagram corpus, rooted at `research-paper-diagram-generation-corpus/`, as builder-time evidence. No new web download was needed for this generation because the local PDF index and extraction artifacts already existed.\n\n## Processing Scope\n\n- Scope: `all_accessible_relevant_pdfs`.\n- Candidate PDFs: 7,631.\n- Accessible PDFs: 7,631.\n- Processed PDFs: 7,631.\n- Skipped PDFs: 0.\n- Skipped reasons: none recorded.\n- Representative rendered pages: 96 audit aids, not the corpus size.\n\n## Framework Evidence Focus\n\nThe generated skill narrows the broader diagram taxonomy to framework-related figure production: method frameworks, architecture diagrams, pipeline/process diagrams, agent workflows, system/data-flow diagrams, mechanism-intuition figures, and case walkthroughs.\n\nProduction-grade lock is supported because the skill includes source-corpus notes, an evidence-map index, an evidence-lineage summary, and framework-specific taxonomy/pattern references.\n\nFile v1.2.0:references/evidence-lineage-summary.md\n\n# Evidence Lineage Summary\n\nThe full local evidence map supports these skill claims:\n\n1. Framework-diagram routing must distinguish method framework, architecture, pipeline/process, agent workflow, system/data flow, graph/network, mechanism, walkthrough, evidence-board, taxonomy, data/protocol, failure, and theory/proof-intuition layouts.\n2. Reader questions differ across diagrams: system identity, process sequence, entity relation, mechanism, case behavior, contribution boundary, and claim support.\n3. Density must be selected before rendering because framework figures range from clean overview panels to dense evidence-linked boards.\n4. Panel choreography is a first-class design decision for multi-panel research figures.\n5. Evidence diagrams must not invent results; they require user-provided metrics, comparisons, or qualitative examples.\n6. Routing must be multi-label: one PDF or one diagram can support multiple labels before a primary production subtype is selected.\n7. When multiple schemes are plausible, the skill should move toward generated candidate images or schematic candidates, usually 6, instead of asking the user to compare only text.\n\nThese claims are backed by `research-paper-diagram-generation-corpus/extracted/evidence_map.json`.\n\n## Full-Corpus Counts\n\n- Processed local PDF records: 7,631.\n- Verified official oral PDFs: 3,356.\n- Supplemental local PDFs: 4,275.\n- Extracted figure captions: 146,071.\n- Diagram-relevant captions: 119,534.\n- Multi-label diagram records: 93,088.\n- Representative rendered pages: 96.\n\n## Framework-Relevant Label Coverage\n\nThe label counts are multi-label and can sum above the number of diagram-relevant captions.\n\n| Label | Caption count | Paper count | Representative pages |\n|---|---:|---:|---:|\n| method_framework | 16,565 | 5,267 | 8 |\n| architecture | 14,651 | 4,256 | 8 |\n| pipeline_process | 28,981 | 5,878 | 8 |\n| agent_workflow | 16,271 | 2,228 | 8 |\n| graph_network | 26,059 | 4,335 | 8 |\n| mechanism_intuition | 42,578 | 5,947 | 8 |\n| case_walkthrough | 35,822 | 6,465 | 8 |\n| evidence_board | 63,404 | 6,947 | 8 |\n| data_benchmark_protocol | 25,583 | 4,881 | 8 |\n| failure_limitation | 20,719 | 4,570 | 8 |\n| taxonomy_design_space | 5,993 | 2,141 | 8 |\n| theory_proof_intuition | 9,985 | 3,286 | 8 |\n| general_diagram_or_figure | 17,072 | 3,160 | 0 |\n\nFile v1.2.0:references/evidence-map-index.md\n\n# Evidence Map Index\n\nThis skill is generated from `research-paper-figure-skill-factory` v1.0.1 using the existing project-local diagram evidence corpus.\n\nPrimary evidence map:\n\n`research-paper-diagram-generation-corpus/extracted/evidence_map.json`\n\nSupporting artifacts:\n\n- `research-paper-diagram-generation-corpus/extracted/paper_cards.csv`\n- `research-paper-diagram-generation-corpus/extracted/paper_cards.json`\n- `research-paper-diagram-generation-corpus/extracted/figure_inventory.csv`\n- `research-paper-diagram-generation-corpus/extracted/figure_inventory.json`\n- `research-paper-diagram-generation-corpus/extracted/caption_inventory.csv`\n- `research-paper-diagram-generation-corpus/extracted/caption_inventory.json`\n- `research-paper-diagram-generation-corpus/extracted/label_summary.csv`\n- `research-paper-diagram-generation-corpus/extracted/panel_structure_notes.md`\n- `research-paper-diagram-generation-corpus/extracted/visual_pattern_observations.md`\n- `research-paper-diagram-generation-corpus/extracted/extraction_report.md`\n- `research-paper-diagram-generation-corpus/extracted/extraction_summary.json`\n- `research-paper-diagram-generation-corpus/extracted/representative_rendered_pages/`\n- `research-paper-diagram-generation-corpus/extracted/diagram_evidence_contact_sheet.jpg`\n- `research-paper-diagram-generation-corpus/metadata/retrieval_manifest.json`\n- `research-paper-diagram-generation-corpus/metadata/acquisition_report.md`\n- `references/builder-time-acquisition-report.md`\n\nUse `label_summary.csv` before routing a new target. The labels are multi-label, so a single paper or diagram may count under multiple classes.\n\nRepresentative rendered pages and contact sheets are audit aids only. Use `retrieval_manifest.json`, `figure_inventory.json`, `caption_inventory.json`, `label_summary.csv`, and `evidence_map.json` as the evidence base.\n\nFile v1.2.0:references/figure-class-taxonomy.md\n\n# Framework Figure Taxonomy\n\nUse this taxonomy before writing prompts or generating images. Routing is multi-label first, then primary subtype selection.\n\n## Core Subtypes\n\n| Subtype | Reader Question | Best Paper Slot | Required Decisions |\n|---|---|---|---|\n| Method framework | What is the proposed method, and why are its parts organized this way? | intro / method | modules, novelty highlight, data/control flow, output |\n| Architecture | Which components interact, and what is trained or inferred? | method / system | boundaries, interfaces, parameters, losses, inference path |\n| Pipeline / process | What happens step by step? | method / system | temporal order, stages, state updates, feedback loops |\n| Agent workflow | What does the agent observe, decide, call, verify, and update? | method / agent system | planner, model/tool calls, memory, verifier, loop |\n| System/data flow | Where do data, users, tools, and model components move? | system / method | lanes, data stores, services, latency or control boundaries |\n| Mechanism intuition | Why does the central idea work? | intro / method / analysis | cause-effect chain, variable roles, constraints |\n| Case walkthrough | How does one example move through the framework? | intro / qualitative / appendix | example states, before/after, stage labels |\n| Evidence-linked framework | How does the framework connect to evidence or ablations? | results / rebuttal | evidence cards, comparison boundary, claim mapping |\n| Failure-aware framework | Where can the system fail, and what boundary should reviewers understand? | analysis / limitation / rebuttal | failure modes, triggers, affected modules, mitigation |\n\n## Routing Axes\n\n- Reader question: identity, sequence, interaction, mechanism, example behavior, evidence support.\n- Logical gap: problem-to-method, method-to-mechanism, mechanism-to-result, result-to-claim.\n- Layout skeleton: left-to-right pipeline, layered stack, hub-and-spoke architecture, swimlanes, modular grid, loop, before/after split.\n- Density: intro overview, method technical, appendix dense, rebuttal conservative.\n- Paper slot: intro, method, system, analysis, appendix, rebuttal, slides.\n- Multi-label status: record all applicable labels, then select a primary subtype.\n\n## Default Primary-Subtype Rule\n\n- For method sections, default to `method_framework` plus `architecture` or `pipeline_process`.\n- For agentic systems, default to `agent_workflow` plus `system_data_flow`.\n- For intro figures, default to `method_framework` plus `mechanism_intuition`.\n- For qualitative examples, default to `case_walkthrough` plus `pipeline_process`.\n- For rebuttals, default to `evidence-linked framework` or `failure-aware framework`.\n\nDo not lock the primary subtype from prose alone when visual comparison would help. Recommend a 6-image candidate board.\n\nFile v1.2.0:references/figure-pattern-library.md\n\n# Framework Figure Pattern Library\n\nUse these layout patterns after the reader effect and primary subtype are known.\n\n## Layout Patterns\n\n| Pattern | Use When | Strength | Risk |\n|---|---|---|---|\n| Left-to-right pipeline | The method is staged and sequential | Fast 10-second comprehension | Can hide feedback loops |\n| Layered architecture stack | Components have abstraction levels | Clear module boundaries | Can feel generic |\n| Hub-and-spoke model center | One core module coordinates inputs/tools/memory | Highlights novelty | Can over-center one block |\n| Swimlane system diagram | Data/user/tool/model roles must be separated | Good for agent/system papers | Needs strong label discipline |\n| Loop / agent workflow | Planning, tool use, verification, or memory updates repeat | Shows dynamics | Arrows can become cluttered |\n| Modular tile board | Many modules or evidence cards need comparison | Scannable, modern | Weak sequence unless arrows are clear |\n| Mechanism snapshot | One mechanism explains why the method works | Strong for intro/method bridge | May omit operational details |\n| Case walkthrough strip | A concrete example travels through the method | Intuitive for qualitative papers | Not enough for full method spec |\n| Baseline-vs-ours split | Novelty is comparative | Reviewer-friendly contrast | Can become adversarial or oversimplified |\n\n## Framework-Specific Prompt Rules\n\n- Give every module a short, exact label.\n- Make novelty visually salient without turning every box into a highlight.\n- Use arrows with semantics: data flow, control flow, feedback, supervision, or evidence link.\n- Keep secondary details in caption/body text unless they are needed for the 10-second reader effect.\n- When multiple layout patterns are plausible, present 4-6 text candidates and recommend a 6-image schematic board.\n\n## Anti-Patterns\n\n- Generic boxes with no novelty hierarchy.\n- Too many equal-weight arrows.\n- Text paragraphs inside the figure.\n- Fake plots or invented numeric results.\n- Decorative icons that do not carry method meaning.\n- Photorealistic clutter for reviewer-sensitive method figures.\n\nFile v1.2.0:references/initial-corpus-manifest.md\n\n# Initial Corpus Manifest Summary\n\nThis package records a summary manifest instead of bundling the full local PDF corpus.\n\n## Manifest Source\n\n- Source corpus root used during generation: `research-paper-diagram-generation-corpus/`\n- Retrieval manifest source: `research-paper-diagram-generation-corpus/metadata/retrieval_manifest.json`\n- Evidence map source: `research-paper-diagram-generation-corpus/extracted/evidence_map.json`\n- Label summary source: `research-paper-diagram-generation-corpus/extracted/label_summary.csv`\n\n## Counts\n\n- Total local PDF records: 7,631.\n- Verified official oral PDFs: 3,356.\n- Supplemental local PDFs: 4,275.\n- Processed PDFs: 7,631.\n- Skipped PDFs: 0.\n- Figure captions: 146,071.\n- Diagram-relevant captions: 119,534.\n- Multi-label figure records: 93,088.\n\n## Use In This Skill\n\nThe skill does not need the full corpus to run on a user's target paper. It uses the corpus-derived taxonomy, pattern library, prompt rules, and review rubric embedded in this package. If the user asks to audit evidence provenance, cite this manifest summary and `references/source-corpus-notes.md`.\n\nFile v1.2.0:references/prompt-generation-policy.md\n\n# Prompt Generation Policy\n\nVersion: 1.2.0\n\n## Prompt Types\n\nThis skill uses two prompt stages:\n\n- **P4 candidate-board brief:** low-commitment board for choosing a direction after text candidates.\n- **P7 final image brief:** formal prompt for the selected direction after P6 candidate review.\n\nDo not use P7 as a substitute for P4/P5/P6. After 4-6 text candidates, the candidate-board bridge must happen unless explicitly skipped by the user.\n\n## Candidate-Board Brief\n\n```text\nBoard purpose:\nGenerate candidate images or schematic candidates for choosing a framework-figure direction, not a final figure.\n\nCandidate count:\n6 by default, allowed 4-6.\n\nHold fixed:\n<paper thesis, target slot, required modules, exact labels, color semantics, sample-image transfer rules>\n\nVary only:\n<subtype / scheme / layout / style / metaphor / density / prompt framing>\n\nCompare:\n<what the user should decide by looking at the images>\n\nRendering route:\nChatGPT web: Create image through ChatGPT Images 2.0.\nCodex: $imagegen first; if unavailable, ChatGPT Images 2.0 API or another approved image-generation API.\n```\n\n## Final Image Brief\n\n```text\nCreate a publication-ready research-paper framework diagram as a raster image.\n\nGoal:\n<figure thesis>\n\nPaper slot and audience:\n<slot and audience>\n\nDiagram subtype and layout:\n<subtype, canvas, panel count, reading order>\n\nRequired content:\n<modules/entities/stages/evidence>\n\nLabels:\nUse only these exact labels: <labels>.\n\nSample-image transfer:\n<per-image transfer rules or \"none\">\n\nStyle and color semantics:\n<style, palette, what colors mean>\n\nCandidate variation:\nGenerate <4/5/6> candidates, usually 6. Vary only <axis>.\n\nAvoid:\nlong paragraphs, microscopic labels, fake metrics, fake UI, logos, watermarks, decorative clutter, SVG/Mermaid/TikZ/Graphviz/code-rendered diagram instructions.\n```\n\n## Rendering Route\n\n- ChatGPT web: use Create image through ChatGPT Images 2.0.\n- Codex: use `$imagegen` first.\n- Codex fallback: ChatGPT Images 2.0 API or another approved image-generation API.\n\nAllowed bitmap outputs: PNG, JPG, JPEG, WebP.\n\nForbidden visual outputs/fallbacks: SVG, Mermaid, TikZ, Graphviz, HTML/CSS, canvas, matplotlib, filesystem code drawing, or code-rendered/exported images.\n\nFile v1.2.0:references/review-rubric.md\n\n# Review Rubric\n\nVersion: 1.2.0\n\nCheck generated framework diagrams against:\n\n- thesis clarity: can a reader understand the main point in 10 seconds?\n- multi-label fit: were all applicable diagram labels considered before selecting the primary subtype?\n- subtype fit: does the layout match the intended primary diagram role?\n- paper fidelity: no invented modules, labels, metrics, or claims;\n- hierarchy: proposed contribution is visually dominant;\n- reading path: arrows and panels have a clear order;\n- label quality: short, readable, and exact;\n- density: enough detail for the paper slot without clutter;\n- sample-image transfer: borrowed only requested attributes from each sample image;\n- candidate-image bridge: text candidates were followed by P4 setup, P5 image-only candidate board, and P6 candidate review before final prompt;\n- rendering route: ChatGPT web used Create image / ChatGPT Images 2.0; Codex used `$imagegen` first or an approved API fallback;\n- rendering safety: no watermark, fake UI, malformed text, or decorative clutter;\n- response boundary: first trigger did not generate images, text turns did not append image generation, and IMAGE_ONLY turns had no prose/state;\n- state footer: every text turn listed all steps, current position, current artifacts, pending artifacts, and next prompts.\n\n## Revision Routes\n\n- If structure is wrong: revise layout skeleton before style.\n- If text is wrong: reduce labels and specify exact text.\n- If novelty is buried: increase contrast and visual weight on the proposed component.\n- If too decorative: switch to formal architecture schematic or minimal line-art.\n- If too sparse: add evidence cards or callouts, but only from provided paper material.\n- If no candidate-board step occurred after text candidates: return to P4 and set up the board before final prompt/final generation.\n\nFile v1.2.0:references/source-corpus-notes.md\n\n# Source Corpus Notes\n\nThis skill was regenerated from the project-local full-feasible diagram corpus used by `research-paper-figure-skill-factory` v1.0.1.\n\nThe underlying extraction artifacts are summarized in this package rather than bundled as thousands of PDFs. The builder run used the project-local corpus rooted at `research-paper-diagram-generation-corpus/`.\n\n## Coverage Summary\n\n- Corpus processing scope: `all_accessible_relevant_pdfs`.\n- Candidate PDF count: 7,631.\n- Accessible PDF count: 7,631.\n- Processed PDF count: 7,631.\n- Skipped PDF count: 0.\n- Skipped reasons: none recorded.\n- Verified official oral PDFs: 3,356.\n- Supplemental local PDFs: 4,275.\n- Figure captions extracted: 146,071.\n- Diagram-relevant captions: 119,534.\n- Multi-label figure records: 93,088.\n- Representative rendered pages: 96, audit aids only.\n\n## Framework-Relevant Evidence\n\n| Label | Caption Count | Paper Count |\n|---|---:|---:|\n| method_framework | 16,565 | 5,267 |\n| architecture | 14,651 | 4,256 |\n| pipeline_process | 28,981 | 5,878 |\n| agent_workflow | 16,271 | 2,228 |\n| mechanism_intuition | 42,578 | 5,947 |\n| case_walkthrough | 35,822 | 6,465 |\n\n## Classification Policy\n\nClassification is multi-label. A single paper or diagram may simultaneously be a method framework, pipeline, architecture, agent workflow, mechanism diagram, evidence board, or case walkthrough. Do not force exclusive labels during routing. First collect all applicable labels, then choose the primary production subtype for the user's target figure.\n\n## Sufficiency\n\n- Evidence sufficiency level: `full_taxonomy`.\n- Lock grade: `production_grade`.\n- Lock basis: `full_taxonomy`.\n\nKnown limitations:\n\n- Evidence extraction used automated PDF text/caption extraction, title/caption multi-label classification, and representative rendered-page inspection; it is not exhaustive manual annotation of every figure.\n- Supplemental local PDFs support taxonomy breadth but are not counted as verified official oral PDFs.\n- For very narrow venue/domain styling, ask for sample images or refresh with closer domain papers.\n\nArchive v1.0.3: 22 files, 38935 bytes\n\nFiles: agents/openai.yaml (292b), assets/conversation_state.template.json (4582b), assets/figure_brief_template.md (1527b), assets/next_step_navigation.md (5717b), assets/prompt_library.md (11863b), assets/refinement_controls.md (1517b), CHANGELOG.md (2614b), examples/example_opening_turns.md (2816b), publish/cover_and_icon_prompts.md (817b), publish/listing_long.md (1294b), publish/listing_short.md (317b), publish/release_checklist.md (555b), publish/starter_messages.md (1580b), README.md (5234b), references/clawhub_packaging_notes.md (1654b), references/README_CN.md (7904b), references/reviewer_style_taxonomy.md (871b), references/visual_communication_principles.md (1765b), references/workflow_examples.md (8092b), SKILL.md (27144b), templates/user_input_bundle.md (863b), _meta.json (152b)\n\nFile v1.0.3:SKILL.md\n\n---\nname: paper-framework-figure-studio-pro\ndescription: Convert a paper deep-reading report, method description, or model introduction into publication-ready framework figure concepts through a stateful multi-round workflow. Use when the user wants top-conference/top-journal framework diagrams, style exploration, human approval between rounds, concrete image prompts, separate text-versus-image turns, and iterative refinement toward a final paper figure.\nlicense: MIT-0\nmetadata:\n  display_name: Paper Framework Figure Studio Pro\n  version: \"1.4.3\"\n  author: OpenAI\n  tags: research-figure, scientific-illustration, framework-diagram, paper-writing, clawhub\n  compatibility: ChatGPT web, Codex, Trae, OpenClaw, ClawHub marketplace, skills.sh. Tool-agnostic. Requires an image-generation capability in the host environment for rendering.\n  openclaw:\n    skillKey: paper-framework-figure-studio-pro\n---\n\n# Paper Framework Figure Studio Pro\nTurn a paper's deep-reading report, model introduction, or method summary into a publication-ready **framework figure** through a **stateful, multi-round, human-in-the-loop studio workflow**.\n\nUse this skill when the user wants any of the following:\n\n- a main paper framework diagram\n- a method overview figure for a conference or journal paper\n- multi-style exploration of academic figure directions\n- iterative narrowing from broad style families to polished final renders\n- concrete, detailed prompts for image generation rather than vague design suggestions\n- explicit human confirmation between rounds\n- a final optional pass that writes the figure legend / caption / panel callouts\n\nThis skill is primarily for **framework figures**, not benchmark plots, tables, or raw quantitative charts.\n\n## Core operating model\nThis skill is not a one-shot prompt generator. It is a **conversation-driven figure studio** with explicit state.\n\n### Mandatory turn separation protocol\nEvery generation cycle must follow this order:\n\n1. **Text-only planning / summary turn**\n2. **Text-only confirmation turn** asking whether to generate the next candidate batch now\n3. **Image-only generation action/turn** after the user says yes\n4. **Text-only evaluation turn** asking the user to choose from the generated images\n\nIf the host environment tends to auto-generate images together with text, the skill must still behave as if those are forbidden to co-occur, and should explicitly defer image generation to the next turn after confirmation.\n\n## First-contact protocol\nOn the **first planning turn**, before the studio fully starts, the assistant should briefly orient the user.\n\n### Preferred but not mandatory upstream input\nThe assistant should explicitly tell the user that the best upstream input is usually a paper deep-reading report in Markdown, and should recommend the user first use the **paper-deep-reading** skill for the target paper or draft when available.\n\nRecommended reminder wording should communicate all of the following:\n\n- best practice is to first generate a paper deep-reading report or structured reading report\n- the report should preferably be saved in **Markdown**\n- the user may use the paper-deep-reading skill at `https://clawhub.ai/c-narcissus/paper-deep-reading` if they want a strong upstream report\n- this is **recommended, not required**\n- the skill also accepts less-complete inputs such as a method sketch, module list, algorithm description, or early-stage design notes\n\n### First-turn readiness check\nAfter the recommendation, the assistant should explicitly ask whether the user is **ready to start figure design now**.\n\nThe first-turn planning reply should therefore do four things in order:\n\n1. remind the user that a Markdown deep-reading report is the preferred input, though not mandatory\n2. state what kinds of partial inputs are also acceptable\n3. ask whether the user is ready to begin figure design now\n4. preview that once the report or description is read, the first image round will usually be a **multi-style candidate board** for the user to choose from\n\n### First turn after ingesting the report or method description\nOnce the assistant has read the user's deep-reading report or method description and extracted the Figure Brief, it should explicitly tell the user that the normal next move is to generate **one batch of multiple style directions** for visual comparison.\n\nThat first post-ingest planning turn should:\n\n- summarize the extracted Figure Brief\n- name the first decision as a **style-family decision**\n- explain that the decision should be made by looking at generated candidate images rather than prose only\n- ask whether to generate the first multi-style candidate board now\nThe assistant should:\n\n1. read the user's paper deep-reading report, method description, or model summary\n2. extract a clean **Figure Brief**\n3. open a multi-round workflow\n4. after each round, summarize the current state and ask for one concrete user decision\n5. **separate all text planning from all image generation**\n6. use **image candidates as the actual decision surface** whenever the user is choosing between visual schemes\n7. update the running state after every user choice and every generation batch\n8. progressively narrow style, structure, density, and detail\n9. after the user selects a final direction, ask whether to also draft the **figure caption / legend / panel explanation text**\n10. end every text-planning reply with a short **Next Steps** block so the user always knows what the next one or two actions will be\n11. after every text-planning reply, record the updated session state, including generated deliverables and user selections\n12. remind the user that future turns should explicitly ask to continue with this skill based on the current saved state\n13. remind the user that if they are unsure what to ask next after images are generated, they can simply type **\"接下来做什么\"** (or **\"what should we do next\"**) to receive guided next-step instructions\n\n## Non-negotiable rules\n- **Human approval gate required** between major rounds.\n- **Do not jump straight to a final image** from the initial paper description.\n- **Do not silently change the chosen direction** without telling the user.\n- **Do not mix planning text and image generation in the same reply** when the host supports separate image actions. First do the text reply. Then do image generation as a distinct action.\n- **Treat this as a hard runtime constraint:** if a reply contains explanation, summary, questions, next-step guidance, or confirmation requests, that reply must be **text-only** and must not trigger image generation.\n- **Before every image batch there must be a dedicated confirmation reply** whose sole job is to ask whether the user wants to generate that batch now. The actual image-generation action must happen only after that confirmation, as a separate next action/turn.\n- **Do not use SVG as the primary or fallback rendering path** for framework-figure candidate boards or finals.\n- **For any round where the user must choose among visual schemes, do not ask them to choose only from prose descriptions. Generate visual candidates first, then ask them to choose by looking at the images.**\n- **Before each new generation batch, explicitly ask whether the user wants to generate the next set of candidate images now.**\n- **Before each generation batch, explicitly name the rendering path that will be used in the current host: ChatGPT web should use native Create image via the assistant under Extended Thinking or the strongest available thinking-assisted path; IDE/API hosts should use OpenAI ChatGPT Images 2.0 or a newer supported OpenAI image model.**\n- **Always track state** using a structure equivalent to `assets/conversation_state.template.json`.\n- **On the first turn, recommend but do not require a Markdown deep-reading report created with the paper-deep-reading skill before figure work begins.**\n- **After reading the report or method description for the first time, explicitly propose generating a first multi-style candidate board for selection.**\n- **After every text-only reply, update the running state with the current figure brief, generated deliverables, pending decisions, and the user's recorded preferences.**\n- **At the end of every text-only reply, remind the user that later messages should explicitly ask this skill to continue from the current state.**\n- **At the end of every text-only planning reply, explicitly remind the user that if they do not know how to continue after the next image batch, they can simply type `接下来做什么` to get guided prompt suggestions and next-step help.**\n- **At the end of every text-only reply, explicitly restate the rendering rule for the current host:** ChatGPT web should use the assistant's native image generation under the strongest available thinking-assisted path (prefer Extended Thinking when available) without asking the user to manually switch to Create image; IDE/API hosts must use OpenAI ChatGPT Images 2.0 or a newer supported OpenAI image model; SVG and other vector-code fallbacks are forbidden.\n- **Generate multiple candidates per batch**. Early exploration batches should usually produce 3 to 5 candidates. Later refinement batches should usually produce 2 to 4 candidates.\n- **Make prompts concrete**. Avoid vague instructions like “make it look academic” unless followed by precise layout, hierarchy, color, metaphor, and typography requirements.\n- **Prefer framework-figure clarity over decorative complexity**.\n- **Ask whether the user wants figure text help after the final image direction is chosen**.\n- **At the end of every text-only planning reply, explicitly tell the user what the next one or two steps are, whether the next step is a text decision or an image-generation step, and what kinds of feedback they should be ready to give after the images appear.**\n\n## Host-specific image-generation policy\n\nWhen this skill reaches a generation step, use the host environment's best available **native OpenAI image-generation path**, and keep it separate from the text-planning reply.\n\nThe separation rule is absolute: a planning / explanation / confirmation message and an image-generation action must never be bundled into one assistant reply.\n\n### Hard prohibition\n\n- **Do not use SVG as the rendering path for candidate boards or final framework figures.**\n- **Do not switch to code-drawn vector output as a substitute for native image generation.**\n- **Do not use mermaid, tikz, graphviz, or other vector-code fallbacks for framework-figure rendering rounds.**\n- The intended rendering path is **OpenAI Create image / ChatGPT Images**, specifically **ChatGPT Images 2.0 or a newer supported OpenAI image model if the host exposes one**.\n\n### ChatGPT web\n\n- The preferred interaction is: the user stays inside the normal chat and the assistant triggers the host's native image generation as a separate image action.\n- **Do not ask the user to manually switch tools or manually click a Create image mode first**; the skill should treat image generation as a native follow-up action after the planning reply.\n- Prefer **Extended Thinking** for framework-figure generation when the host exposes it.\n- If the host experience exposes **Thinking** or **images with thinking** but not an explicit Extended Thinking label, prefer the strongest available reasoning-assisted image path.\n- Treat the image step as its own action after the user says to proceed.\n\n### OpenClaw, Codex, Trae, or other IDE / API-driven hosts\n\n- Use **OpenAI ChatGPT Images 2.0** at minimum for raster image output.\n- If the host exposes a newer OpenAI image-generation version than ChatGPT Images 2.0, use the newer supported OpenAI version.\n- If the host requires an API key and no OpenAI API key is available, **pause before generation and explicitly tell the user that image generation cannot proceed until they provide or configure an OpenAI API key**.\n- When generation is blocked by missing credentials, do not fake progress and do not switch to SVG as a fallback.\n- Do **not** replace image generation with SVG, mermaid, tikz, graphviz, or other vector-code fallbacks when the task is a framework-figure rendering step.\n\n### Required interaction split\n\n- Keep the **text planning reply** and the **image-generation action** separate.\n- When a round is a **visual decision round**, the normal sequence is:\n  1. text turn: summarize state, remind the user of the rendering path that will be used in this host, and ask whether to generate the next candidate board\n  2. image action: use **OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation** to generate the candidate board or multi-image batch\n  3. text turn: briefly label the shown candidates and ask the user to choose by image number / letter\n\n## Visual-decision-first protocol\n\nThis skill uses a **visual-decision-first** workflow.\n\nThat means:\n\n- The assistant may explain what varies across options in text.\n- But when the user's next decision is fundamentally about figure appearance, layout, style family, internal visual language, or refinement direction, the assistant should not stop at text-only options.\n- Instead, the assistant should ask whether to generate a **candidate board** for that decision.\n- After generation, the assistant should present a short mapping such as **A / B / C / D** or **1 / 2 / 3 / 4** tied to the generated images and ask the user to choose from the images.\n\nGood visual-decision rounds include:\n\n- style family selection\n- structural skeleton selection\n- density / audience-bias tradeoff when it changes the figure look\n- internal visual language selection\n- refinement direction selection\n- final shortlist selection\n\n\n## Mandatory next-step navigation\n\nEvery **text-only** reply in the workflow must end with a short navigation block.\n\nThe navigation block should be concrete and user-facing, not abstract process language. It should tell the user:\n\n1. what the very next step is\n2. whether that next step is **another text decision** or a **separate image-generation action**\n3. what the user will need to do right after images are generated\n4. what kinds of feedback will be most useful in the next turn\n\nIt should also repeat one short **rendering-rule reminder** in every text-only reply so the user sees it every round:\n\n- In **ChatGPT web**, image generation must be a separate native image action under the strongest available thinking-assisted path; prefer **Extended Thinking** when available, and do **not** ask the user to manually switch to Create image.\n- In **OpenClaw / Codex / Trae / API hosts**, image generation must use **OpenAI ChatGPT Images 2.0** at minimum, or a newer supported OpenAI image model if available.\n- **SVG, mermaid, tikz, graphviz, and other vector-code fallbacks are forbidden.**\n\n### Required structure\n\nUse a compact structure such as:\n\nIn addition to the user-facing navigation block, the assistant should also internally update the saved session state after the text reply is composed. That state update should include:\n\n- the latest accepted inputs\n- the current Figure Brief\n- every generated deliverable so far\n- the most recent user choices and rejected options\n- the currently pending decision\n- the next candidate board that would be generated if the user says yes\n\n\n- **Next step:** [ask permission to generate the next candidate board / summarize a shortlisted direction / write caption text]\n- **After the images appear, please choose by image** and optionally comment on: [layout / density / comparison clarity / icon style / equations / mini-result snapshots / clutter / reviewer-friendliness]\n- **Then I will:** [update the chosen direction and prepare the next narrower batch]\n- **For the next turn:** please explicitly ask `paper-framework-figure-studio-pro` to continue from the current saved state.\n\n### Session continuity reminder\n\nBecause some hosts do not automatically preserve skill-specific working memory in a reliable way, the assistant should remind the user at the end of each text-only planning turn that future messages should explicitly say something like:\n\n- “Please continue with **Paper Framework Figure Studio Pro** from the current saved state.”\n- “Use **paper-framework-figure-studio-pro** to continue from the current state and apply this new change request.”\n\nThis reminder should be brief, but it should appear consistently so the user knows how to resume the workflow in later turns. The assistant should also record the updated session state after every text-planning reply, including generated deliverables and user selections.\n\n### Examples\n\nExample A:\n\n- **Next step:** If you want, the next action is to generate the style-family candidate board as a separate image batch.\n- **After the images appear, please choose by image** (A/B/C/D) and tell me what you like or dislike about layout, modernity, and explanation strength.\n- **Then I will:** update the state and prepare the next structural-skeleton board.\n\nExample B:\n\n- **Next step:** The next action is to generate a refinement batch focused only on reducing clutter and strengthening the baseline-vs-ours comparison.\n- **After the images appear, please choose by image** and note whether you want fewer labels, cleaner arrows, or stronger mini-result snapshots.\n- **Then I will:** lock the winning direction and ask whether you also want caption / legend / panel text.\n\nDo not omit this navigation block. The user should always know the next one or two moves.\n\n## Workflow overview\n\nFollow this sequence unless the user explicitly asks to skip or compress a stage.\n\n### Round 0 — Intake and figure brief construction\n\nRead the user's deep-reading report or model description and construct a **Figure Brief**.\n\nThe Figure Brief must capture at least:\n\n- paper or method title\n- one-sentence scientific claim\n- what the figure must explain\n- target figure type: framework overview\n- likely venue level and audience familiarity\n- mandatory modules to show\n- optional modules to compare\n- what should remain outside the figure\n- preferred page format: A4 portrait, A4 landscape, or unknown\n- whether the user values safety, modernity, mechanism explanation, or visual memorability more\n\nUse the template in `assets/figure_brief_template.md`.\n\n### Round 1 — Style-family candidate board\n\nDo **not** ask the user to choose only from a written list of families.\n\nInstead:\n\n1. summarize 3 to 5 candidate families very briefly in text\n2. ask whether to generate the **style-family candidate board now**\n3. in a separate image action, generate 3 to 5 style-distinct figure candidates for the same paper content\n4. after the images are shown, label them in a short text turn and ask the user to choose a primary direction and optionally a backup\n\nRecommended default families:\n\n1. **Academic Conservative** — standard top-tier ML paper overview\n2. **Modern Modular Tiles** — magnetic-card / dashboard-like figure blocks\n3. **Mechanism + Result Snapshots** — each stage shows both mechanism and local effect\n4. **Editorial Flat Illustration** — modern flat/cartoon academic style, friendly but rigorous\n5. **Premium Scientific Illustration** — soft-3D / high-polish scientific editorial rendering\n\n### Round 2 — Structural-skeleton candidate board\n\nWithin the selected family, do not stop at textual skeleton descriptions.\n\nInstead:\n\n1. propose 2 to 4 structural skeletons very briefly\n2. ask whether to generate the **structural-skeleton candidate board now**\n3. in a separate image action, generate 2 to 4 candidates where the content stays fixed but the composition changes, such as:\n   - left-to-right pipeline\n   - top-down narrative stack\n   - central model + surrounding callouts\n   - modular tile grid\n   - comparison split with baseline vs ours\n4. after the images are shown, ask the user to choose from the images\n\n### Round 3 — Density / reviewer-bias candidate board\n\nIf density or reviewer bias will materially affect the visual appearance, do not ask the user to decide only from prose.\n\nInstead:\n\n1. explain that the next board will compare, for example:\n   - technical / formal\n   - cross-domain / easier to understand\n   - visually modern but still rigorous\n   and/or\n   - low density\n   - medium density\n   - high density\n2. ask whether to generate the **density-and-bias candidate board now**\n3. generate 2 to 4 candidates in a separate image action\n4. ask the user to choose from the images\n\n### Round 4 — Internal visual-language candidate board\n\nNarrow the figure's visual language.\n\nTypical selectable elements:\n\n- avatars or no avatars\n- mini scatterplots or no mini scatterplots\n- per-step result snapshots or mechanism only\n- one equation or several small equations\n- minimal labels or richer callout labels\n- baseline comparison included or deferred\n\nProtocol:\n\n1. summarize what will vary\n2. ask whether to generate the **internal-visual-language board now**\n3. generate 2 to 4 candidates in a separate image action\n4. ask the user to choose from the images\n\n### Round 5 — Exploration batch generation\n\nPrepare a concrete batch with 3 to 5 candidate prompts.\n\nImportant:\n\n- keep the paper content fixed\n- vary only a few style axes per batch\n- state clearly what differs across candidates\n- ask whether to generate this batch now\n- after the user approves, perform image generation in a separate action\n- after the images appear, ask the user to choose from the actual images rather than from abstract prose\n\nThen update state with the generated batch metadata.\n\n### Round 6 — Selection and refinement\n\nAfter the user chooses a winner or shortlist:\n\n- summarize what won\n- summarize what the user disliked\n- propose the next refinement axis\n- ask whether to generate the narrower refinement batch now\n- generate the refinement batch in a separate image action\n- after the images appear, ask the user to choose from the actual images\n\nTypical refinement axes:\n\n- stronger hierarchy\n- less clutter\n- more legible equations\n- better baseline-vs-ours comparison\n- cleaner client graph\n- more journal-like typography\n- more modern or less playful icons\n- closer to A4 publication balance\n\n### Round 7 — Finalization\n\nOnce the user selects a final direction:\n\n- confirm the final figure intent\n- ask whether they also want:\n  - panel labels\n  - legend text\n  - figure caption\n  - figure explanation for the paper body\n  - bilingual callout wording\n\n## Mandatory text-turn protocol\n\nEvery non-image reply should follow this pattern.\n\n### A. Current state\n\nBriefly state:\n\n- current round\n- current chosen family and skeleton\n- current unresolved decision\n\n### B. Visual decision to be made\n\nIf the next decision is visual, say that the next step should be based on **candidate images**, not only verbal descriptions.\n\n### C. Ask permission for the next image batch\n\nAsk a bounded confirmation such as:\n\n- “Do you want me to generate the next style-family candidate board now?”\n- “Do you want me to generate the structural-layout candidates now?”\n- “Do you want me to generate the next refinement batch now?”\n\n### D. What the next batch will vary\n\nState 2 to 5 controlled axes that will differ across the generated images.\n\n### E. After images are shown\n\nIn the next text turn after generation:\n\n- label the shown candidates clearly\n- give a one-line difference summary for each candidate\n- ask the user to choose by image ID, for example **A**, **B**, **C**, **D**\n\n## Mandatory image-turn protocol\n\nEvery image-generation step must be independent from the planning text turn.\n\n- Do not include a long discussion inside the image-generation step.\n- The generation action should use a concrete prompt assembled from the current state.\n- Early rounds should produce multiple style-diverse candidates.\n- Later rounds should produce tightly controlled refinements.\n- The batch should be assembled so that the user can make a real choice **from the generated images**.\n\n## Prompt-construction standard\n\nBuild prompts from the following layers, in this order.\n\n1. **Figure goal** — what the figure explains scientifically\n2. **Paper framing** — title and one-line claim\n3. **Required content blocks** — the exact modules that must appear\n4. **Narrative order** — the intended reading path\n5. **Style family** — one of the chosen families\n6. **Structural skeleton** — layout archetype\n7. **Visual vocabulary** — icons, nodes, mini plots, avatars, tiles, cards\n8. **Typography requirements** — concise labels, sharp text, panel headings\n9. **Color semantics** — blue shared, orange personal, green collaboration by default\n10. **Comparative emphasis** — consensus-only vs beyond consensus if included\n11. **Output constraints** — A4, portrait/landscape, publication-ready, uncluttered, legible\n12. **Batch-difference instruction** — what should differ across candidate A/B/C/D and what must stay fixed\n\nUse the detailed templates in `assets/prompt_library.md`.\n\n## Visual-communication standards\n\nFramework figures should satisfy the following principles.\n\n- one dominant message per figure\n- strong reading path\n- consistent visual metaphor\n- stable color semantics across rounds\n- enough white space to separate reasoning chunks\n- panel labels must reflect conceptual boundaries, not arbitrary boxes\n- if mini result snapshots are used, they must illustrate a real conceptual change rather than act as decoration\n- decorative flair must never obscure the core method\n- image batches should differ along deliberate axes that are visible enough for the user to judge from the images\n\n---\nSee `references/visual_communication_principles.md`.\n\n## Common failure modes to avoid\n\n- asking the user to choose a visual direction from text only when images are required to judge it\n- forgetting to ask whether to generate the next candidate board now\n- mixing the explanation turn and the image turn into a single blended reply\n- too much tiny text inside the image\n- mixing too many styles in one batch\n- icons that imply the wrong algorithmic semantics\n- confusing “shared” with “global final model” when the paper is personalized\n- making every step equally visually heavy\n- comparison panel larger than the main mechanism\n- decorative 3D effects that damage legibility\n- using unrealistic benchmark plots when the figure is supposed to be a framework diagram\n- asking the user too many open-ended questions at once\n\n## What to ask after a final figure is chosen\n\nAlways ask:\n\n- Do you want me to also write the **figure caption**?\n- Do you want **panel-wise explanatory text** for the paper body or appendix?\n- Do you want a **short legend / callout wording pass** to improve what appears inside the figure?\n\n## Files in this skill bundle\n\n- `assets/figure_brief_template.md`\n- `assets/conversation_state.template.json`\n- `assets/prompt_library.md`\n- `assets/refinement_controls.md`\n- `references/visual_communication_principles.md`\n- `references/reviewer_style_taxonomy.md`\n- `references/workflow_examples.md`\n- `references/README_CN.md`\n\nUse them actively rather than improvising from scratch every round.\n\nFile v1.0.3:README.md\n\n# Paper Framework Figure Studio Pro\n\n**Paper Framework Figure Studio Pro** is a stateful, multi-round scientific-figure skill for turning a paper deep-reading report, method summary, module sketch, or algorithm description into a publication-ready **framework figure workflow**.\n\nIt is designed for **top-tier CS paper figures** where users want:\n\n- explicit human confirmation between rounds\n- visual candidate boards before style/layout decisions\n- text planning and image generation kept separate\n- OpenAI native image generation only for figure renders\n- recorded state across rounds\n- final optional help with caption, legend, and panel explanation text\n\nThis package is prepared as a **publish-ready skill bundle** for OpenClaw / ClawHub style runtimes and similar hosts.\n\n## What this skill does\n\nThe skill reads the user's paper or method description, builds a **Figure Brief**, then runs a guided studio workflow:\n\n1. Recommend a Markdown deep-reading report as the best upstream input (but do not require it)\n2. Confirm the user is ready to begin figure design\n3. Extract the Figure Brief\n4. Propose the first **multi-style candidate board**\n5. Generate multiple image candidates as a **separate image action**\n6. Ask the user to choose by looking at the generated images\n7. Update state, narrow direction, and continue to the next refinement round\n8. Ask at the end whether the user also wants caption / legend / panel explanation support\n\n## Hard rules\n\n- **No SVG rendering path** for candidate boards or final framework figures\n- **No mermaid / graphviz / tikz fallback** for figure-rendering rounds\n- Text planning and image generation must be **separate steps**\n- **A reply may never contain both planning text and image generation.** If the assistant is asking, explaining, summarizing, or requesting confirmation, that reply must be text-only.\n- **Before each generation batch, there must be a dedicated text-only confirmation turn** asking whether to generate the next candidate images now. The actual image generation must happen only in the next separate action/turn after the user confirms.\n- Visual decisions should be made from **generated images**, not prose-only descriptions\n- Every text turn must end with **Next Steps** guidance\n- Every text turn must update and preserve session state\n\n## Image-generation policy\n\n### ChatGPT web\nUse the host's native **Create image** path as a separate action after the planning reply. Prefer **Extended Thinking** or the strongest available thinking-assisted image path exposed by the host. Do not instruct the user to manually switch tools first.\n\n### OpenClaw / Codex / Trae / IDE / API hosts\nUse **OpenAI ChatGPT Images 2.0** at minimum, or a newer supported OpenAI image model if exposed by the host. If no OpenAI API key is available, the skill must pause and ask the user to provide or configure one before generation.\n\n## Package contents\n\n- `SKILL.md` — main skill specification\n- `LICENSE` — MIT-0 / MIT No Attribution license text\n- `VERSION` — current package version\n- `CHANGELOG.md` — release notes\n- `assets/` — working templates and navigation / prompt libraries\n- `references/` — Chinese guide, workflow notes, reviewer taxonomy, visual communication principles\n- `examples/` — suggested opening turns and continuation patterns\n- `publish/` — release-page copy, listing text, icon / cover prompts, publishing checklist\n- `templates/` — optional user-facing input templates\n\n## Suggested release metadata\n\n- **Slug:** `paper-framework-figure-studio-pro`\n- **Name:** `Paper Framework Figure Studio Pro`\n- **License:** MIT-0\n\n## Quick start\n\nThe best first user message is something like:\n\n> Please use **paper-framework-figure-studio-pro**. I have a deep-reading report in Markdown for my paper draft. Read it, extract the figure brief, and tell me whether we should generate the first multi-style candidate board.\n\nOr, for an early-stage project:\n\n> Please use **paper-framework-figure-studio-pro**. I do not have a full draft yet. I only have a model description and module design notes. Read them, build the figure brief, and tell me whether I am ready to start the first candidate board.\n\n## Recommended upstream companion skill\n\nThis skill works best when the user first prepares a paper reading report in Markdown with **paper-deep-reading**. The skill should recommend, but not require, the upstream deep-reading workflow.\n\n## Release note\n\nThis package is intentionally focused on **framework figures**. It is not a plot generator, chart generator, or general slide-design skill.\n\n- At the end of every text-only planning reply, the studio should remind the user that if they are unsure how to continue after the next image batch, they can simply type **`接下来做什么`** to receive guided next-step instructions.\n\n\nRendering rule reminder used in every text-only reply:\n- ChatGPT web: use native image generation under the strongest available thinking-assisted path; prefer Extended Thinking when available; do not ask the user to manually switch to Create image.\n- OpenClaw / Codex / Trae / API hosts: use OpenAI ChatGPT Images 2.0 or newer.\n- SVG, mermaid, tikz, graphviz, and other vector-code fallbacks are forbidden.\n\nFile v1.0.3:_meta.json\n\n{\n  \"ownerId\": \"kn7fxns1xpr6z67w885my7d7k98506vv\",\n  \"slug\": \"paper-framework-figure-studio-pro\",\n  \"version\": \"1.0.3\",\n  \"publishedAt\": 1776945905161\n}\n\nFile v1.0.3:references/clawhub_packaging_notes.md\n\n# ClawHub Packaging Notes\n\nThis bundle is prepared for ClawHub / OpenClaw style packaging:\n\n- directory name matches the publish slug: `paper-framework-figure-studio-pro`\n- `SKILL.md` uses YAML frontmatter\n- the frontmatter `name` is the lowercase publish-safe skill identifier\n- the human-readable display name is stored in `metadata.display_name`\n- the license is declared as `MIT-0` and the full MIT No Attribution text is included in `LICENSE`\n- the package version is recorded in both `SKILL.md` metadata and `VERSION`\n- OpenClaw runtime metadata is declared under `metadata.openclaw`\n- no conflicting license terms are added inside `SKILL.md`\n\nIf a publish UI asks separately for display name, use:\n\n- **Slug**: `paper-framework-figure-studio-pro`\n- **Name / Display Name**: `Paper Framework Figure Studio Pro`\n\n\n## Image-rendering constraint\n\nThis skill is intentionally authored for **native image generation** rather than SVG synthesis. In hosts that support separate image actions, the recommended path is a distinct **Create image** step, ideally using **Thinking** or **Extended Thinking / images with thinking** when available for complex framework figures.\n\n\nAdditional host policy for v1.2.1:\n- ChatGPT web: keep the user in chat, prefer Extended Thinking or the strongest available thinking-assisted image path, and do not ask the user to manually switch tools before generation.\n- IDE / API hosts such as OpenClaw, Codex, and Trae: use OpenAI ChatGPT Images 2.0 at minimum; if credentials are missing, stop and ask the user to configure an OpenAI API key before any generation round.\n- Never downgrade framework-figure rendering to SVG.\n\nFile v1.0.3:references/README_CN.md\n\n# 发布版说明\n\n当前目录已经补齐为一个更完整的发布版 skill 包，额外包含：\n\n- `README.md`：英文发布页说明\n- `LICENSE`：MIT-0 正式授权文本\n- `CHANGELOG.md`：版本记录\n- `examples/`：示例开场与续接写法\n- `publish/`：ClawHub / OpenClaw 上架页文案、封面与图标 prompt、发布检查清单\n- `templates/`：用户输入模板\n\n---\n\n# Paper Framework Figure Studio Pro 使用说明（中文）\n\n这是一个面向论文**框架图**的多轮科研绘图 skill。\n\n它不是“直接给一条提示词然后开画”，而是：\n\n1. 先读取用户的论文精读报告、方法介绍或模型说明\n2. 抽取 Figure Brief\n3. 开启多轮对话\n4. 每一轮只让用户做一个关键决定\n5. **凡是涉及视觉方案选择，先独立生出候选图，再让用户看图选择**\n6. 每次生图都单独执行，不和解释文字混在同一回复里\n7. 用户选图后更新状态，再进入下一轮细化\n8. 最终询问用户是否还要补充图注、caption、图中文字说明\n\n## 这版 skill 特别强调的协议\n\n### 1. 文字步骤和生图步骤严格分开\n\n正确节奏应该是：\n\n- 第一步：文字回复，说明当前状态，并询问“是否现在生成下一轮候选图”\n- 第二步：独立执行生图动作（Create image / 对应 API）\n- 第三步：文字回复，对已经生成的候选图做编号，然后请用户**根据图来选**\n\n不要把“解释 + 生图 + 让用户选择”揉成一条混合回复。\n\n### 1.5 生图路径必须走 OpenAI Create image / ChatGPT Images 2.0，而不是 SVG\n\n- 框架图候选图和最终图，**不要走 SVG 合成路线**。\n- 在 **ChatGPT web** 中，应走独立的 **Create image** 步骤。\n- 若宿主支持 **Thinking** 或 **Extended Thinking / images with thinking**，优先用这一路径来生成复杂框架图。\n- 在 **Codex / Trae / API 宿主** 中，也应走原生的 ChatGPT 图片生成能力；不要把 mermaid、graphviz、tikz、纯 SVG 输出当成框架图渲染替代方案。\n\n### 2. 让用户选方案时，应优先让用户从图里选\n\n不是：\n\n- 先写一堆方案文字说明\n- 然后让用户凭想象选 A / B / C\n\n而应该是：\n\n- 先简要说明下一批图会比较什么\n- 询问用户是否现在生成这一批候选图\n- 生成多张图\n- 再让用户从图里选择 A / B / C / D\n\n\n### 4. 每次文字回复后，都要告诉用户下面 1–2 步做什么\n\n不要只停在当前轮的说明上。每次文字回复结束时，都应该显式告诉用户：\n\n- 下一步是不是要单独生成下一批图\n- 生成完后用户应该从哪些维度来选图\n- 用户选完之后，再下一步会进入哪一轮细化\n\n推荐固定加一个小结尾：\n\n- **下一步**：是否现在生成这一轮候选图\n- **看图后请你重点反馈**：例如风格、结构、密度、机制解释强度、是否太花、是否太满\n- **然后我会**：更新状态并进入下一轮更细的候选图\n\n### 3. 每一轮开始前都可以问一句\n\n推荐问法：\n\n- “要不要我先生成这一轮的风格候选图？”\n- “要不要我先生成这一轮的结构候选图？”\n- “要不要我先生成下一轮细化图？”\n\n## 推荐的多轮顺序\n\n- 第 0 轮：读取论文内容，整理 Figure Brief\n- 第 1 轮：先选大风格家族，但应通过**风格候选图**来选\n- 第 2 轮：再选结构骨架，但应通过**结构候选图**来选\n- 第 3 轮：再选面向哪类审稿人 + 信息密度，但如果视觉差异明显，也应通过**候选图**来选\n- 第 4 轮：再选人物图标、结果示意、公式多少、对比区大小，并通过**内部视觉语言候选图**来选\n- 第 5 轮：首批综合方案多图生成\n- 第 6 轮：用户选图，归纳喜欢和不喜欢的点\n- 第 7 轮：第二批定向细化生成\n- 第 8 轮：最终定稿 + 询问是否补 caption / legend / panel explanation\n\n## 这个 skill 特别强调的点\n\n- 每次生图前都要有人类确认\n- 每次生成要多张图，不要一开始只出 1 张\n- 提示词要具体，不能只写“学术风”“顶刊风”\n- 图像生成动作必须和普通文字回复分离\n- 凡是视觉方向决策，优先基于**已生成图**来做选择\n- 每轮结束后都要更新状态\n- 最终一定要问用户是否还需要图注和正文配套说明\n\n\n### 1.8 宿主环境提醒必须明确说明\n\n- 如果运行在 **OpenClaw / Codex / Trae / 其他 IDE 或 API 宿主**，框架图生图阶段必须走 **OpenAI ChatGPT Images 2.0**（若宿主提供更高版本，则用更高版本）。\n- 如果该宿主没有可用的 **OpenAI API key**，必须先提醒用户提供或配置 key，再进入生图步骤。\n- 如果运行在 **ChatGPT 网页版**，应要求在 **Extended Thinking** 或宿主当前可用的最强 thinking-assisted 路径下进行，并且**不要让用户手动切换到 Create image 工具模式**；而是由助手在独立的生图步骤中触发原生图片生成。\n- 无论在哪个宿主里，**都不能因为缺少图片能力而改用 SVG**。\n\n## 首次进入 skill 时的提醒\n\n第一次使用这个 skill 时，应该先提醒用户：\n\n- **最佳做法**：先用 `paper-deep-reading` skill 对论文初稿或相关论文生成一份**精读报告**，并保存为 **Markdown**\n- 推荐链接：`https://clawhub.ai/c-narcissus/paper-deep-reading`\n- 但这**不是必须的**\n- 如果用户还没有准备好初稿，也可以先输入：模型模块描述、算法思路、设计思想、算法流程、训练机制、系统结构草图等\n\n然后要继续确认两件事：\n\n1. 用户当前输入是否已经足够开始框架图设计\n2. 用户是否准备好**现在开始画图**\n\n一旦读完精读报告或方法描述，第一次正式设计轮应该明确告诉用户：\n\n- 下一步通常是先生成**一批不同风格的候选图**供选择\n- 这个选择应该基于图，而不是只基于文字描述\n\n\n## 状态记录与后续续接\n\n每次**文字回复**后，除了给用户当前结论，还应更新一份当前状态，至少记录：\n\n- 当前 Figure Brief\n- 已经生成了哪些候选图 / 交付物\n- 用户已经选中了什么、排除了什么\n- 当前正在等待哪一个决定\n- 下一步若用户同意，会生成哪一批图\n\n同时还要提醒用户：为了确保后续在 OpenClaw、Codex、Trae 或其他宿主里能稳定续接，后续每次提问时，最好显式写上类似：\n\n- `请使用 paper-framework-figure-studio-pro 根据当前状态继续执行，并处理下面的新要求：...`\n- `Use paper-framework-figure-studio-pro to continue from the current saved state and apply the following change: ...`\n\n这样能减少宿主环境丢失上下文或没有正确续接状态的风险。\n\n\n## 关键硬约束：文字回复和生图不能同轮出现\n\n- 只要这一轮回复里包含解释、总结、追问、确认、下一步导航，就必须是**纯文字回复**。\n- 如果下一步要生图，这一轮只能先说明将要生成什么，并询问用户是否现在开始生成。\n- 用户确认后，**下一轮 / 下一独立动作**才允许真正调用 OpenAI ChatGPT Images 2.0 / Create image。\n- 绝不能在“说明 + 生图”同一次回答里同时发生。\n\n- 每次文字回复结尾都要提醒用户：如果生成图后不知道下一步该怎么提问，可以直接输入 **`接下来做什么`**，skill 会给出下一步建议和推荐提问方式。\n\n\nRendering rule reminder used in every text-only reply:\n- ChatGPT web: use native image generation under the strongest available thinking-assisted path; prefer Extended Thinking when available; do not ask the user to manually switch to Create image.\n- OpenClaw / Codex / Trae / API hosts: use OpenAI ChatGPT Images 2.0 or newer.\n- SVG, mermaid, tikz, graphviz, and other vector-code fallbacks are forbidden.\n\nFile v1.0.3:references/reviewer_style_taxonomy.md\n\n# Reviewer Style Taxonomy\n\nUse this to tailor the figure direction to the likely reviewer.\n\n## Technical-formal reviewer\nPrefers:\n- cleaner architecture diagrams\n- fewer decorative icons\n- stronger notation discipline\n- moderate or high information density\n\nGood families:\n- Academic Conservative\n- Modern Modular Tiles (formal version)\n- Mechanism + Result Snapshots (formal version)\n\n## Cross-domain reviewer\nPrefers:\n- clearer storytelling\n- stronger intuition cues\n- fewer symbols\n- more obvious problem-to-method narrative\n\nGood families:\n- Modern Modular Tiles\n- Editorial Flat Illustration\n- Mechanism + Result Snapshots\n\n## Design-sensitive reviewer\nPrefers:\n- polished hierarchy\n- modern composition\n- memorable visual system\n- elegance without loss of rigor\n\nGood families:\n- Modern Modular Tiles\n- Editorial Flat Illustration\n- Premium Scientific Illustration\n\nFile v1.0.3:references/visual_communication_principles.md\n\n# Visual Communication Principles for Top-Tier Research Framework Figures\n\n## 1. One dominant message\nA strong framework figure should communicate one dominant thesis. For this paper family, the thesis is typically:\n\n> collaborate on what is shareable, preserve what is personal.\n\nEverything in the figure should support that sentence.\n\n## 2. Reading path matters more than decoration\nTop-tier figures usually succeed because the reading path is obvious:\n\n- either left-to-right\n- or top-down\n- or center-with-radial-callouts\n\nA viewer should know where to look first within one second.\n\n## 3. Hierarchy must be visible, not only logical\nImportant modules should be visually larger, cleaner, or more central. Not every panel deserves equal weight.\n\n## 4. Stable color semantics\nIf blue means shared structure once, it should mean shared structure everywhere. Do not reuse colors for conflicting concepts.\n\n## 5. Abstract methods benefit from concrete micro-evidence\nIf the method is conceptually abstract, small micro-plots or state snapshots can significantly improve comprehension.\n\n## 6. Use equations sparingly\nInside the image, one or two compact equations help anchor rigor. Too many equations reduce figure readability.\n\n## 7. Comparison panels should clarify the paper's delta\nA baseline-vs-ours panel should answer: what is qualitatively different here?\n\n## 8. White space is semantic structure\nWhite space is not empty. It separates ideas, controls reading flow, and prevents false grouping.\n\n## 9. Avoid decorative ambiguity\nDo not use fancy icons or 3D elements unless they clarify rather than distract.\n\n## 10. The figure should survive both quick scan and close read\nGood paper figures work at two speeds:\n- 10-second scan\n- 60-second careful reading\n\nFile v1.0.3:references/workflow_examples.md\n\n# Workflow Examples\n\n## Example 0 — First contact and readiness check\n\nRound 0 text:\n- remind the user that the preferred upstream input is a Markdown deep-reading report\n- recommend the paper-deep-reading skill at `https://clawhub.ai/c-narcissus/paper-deep-reading`\n- make clear that this is recommended but not required\n- accept partial inputs such as a method sketch, module description, algorithm notes, or early-stage design ideas\n- ask whether the user is ready to start figure design now\n- explain that once the report or description is read, the first image round will normally be a **multi-style candidate board**\n- end with a Next Steps block and a resume reminder telling the user to later ask this skill to continue from the current state\n\nRound 1 text after ingesting the report or method description:\n- summarize the extracted Figure Brief\n- say that the first decision should be made by **looking at images** rather than prose only\n- ask whether to generate the first multi-style candidate board now\n- end with a Next Steps block and a resume reminder\n\n## Example 1 — Safe conference-paper path\n\nRound 0: ingest report and create Figure Brief\nRound 1 text: summarize style-family options and ask whether to generate the style-family candidate board now\nRound 1 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation (prefer Extended Thinking or the strongest available thinking-assisted path when available) to generate 4 style-family candidates\nRound 1 follow-up text: label A/B/C/D and ask the user to choose by image\nRound 2 text: summarize the chosen family and ask whether to generate the structural-skeleton board now\nRound 2 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 3 structure candidates\nRound 2 follow-up text: ask the user to choose by image\nRound 3 text: ask whether to generate the density/bias board now\nRound 3 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 3 density variants\nRound 3 follow-up text: ask the user to choose by image\nRound 4 text: ask whether to generate the internal-visual-language board now\nRound 4 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 3 visual-language candidates\nRound 4 follow-up text: ask the user to choose by image\nRound 5 image action: generate 4 integrated exploration candidates\nRound 6: user selects candidate B and requests less clutter\nRound 6 text: summarize keep/change list and ask whether to generate the next refinement batch now\nRound 6 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 3 refinements\nRound 7: user selects final and asks for caption\n\n## Example 2 — Mechanism-explanation path\n\nRound 0: ingest abstract method description\nRound 1 text: explain that style choice should be image-based and ask whether to generate the first candidate board now\nRound 1 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation (prefer Extended Thinking or the strongest available thinking-assisted path when available) to generate 3 style candidates with different mechanism-explanation intensities\nRound 1 follow-up text: ask the user to choose by image\nRound 2 text: ask whether to generate top-down vs left-to-right structure candidates now\nRound 2 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 2 structure candidates\nRound 2 follow-up text: ask the user to choose by image\nRound 3 text: ask whether to generate medium vs medium-high density candidates now\nRound 3 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 2 density candidates\nRound 3 follow-up text: ask the user to choose by image\nRound 4 text: ask whether to generate mini-scatterplot and result-snapshot variants now\nRound 4 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 3 variants\nRound 4 follow-up text: ask the user to choose by image\nRound 5: integrated candidate batch\nRound 6: refinement\nRound 7: caption + panel explanation\n\n## Example 3 — Visually memorable flagship path\n\nRound 0: ingest full deep-reading report\nRound 1 text: summarize Premium Scientific Illustration and Editorial Flat Illustration, then ask whether to generate the style board now\nRound 1 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation (prefer Extended Thinking or the strongest available thinking-assisted path when available) to generate 4 flagship-style candidates\nRound 1 follow-up text: ask the user to choose by image and optionally nominate a backup\nRound 2 text: ask whether to generate center-with-callouts vs modular-card structures now\nRound 2 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 2 structural candidates\nRound 2 follow-up text: ask the user to choose by image\nRound 3 text: ask whether to generate medium-density / formal-modern variants now\nRound 3 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 2 variants\nRound 3 follow-up text: ask the user to choose by image\nRound 4 text: ask whether to generate icon-vocabulary candidates now\nRound 4 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation to generate 3 icon-language candidates\nRound 4 follow-up text: ask the user to choose by image\nRound 5: exploration batch\nRound 6: shortlist two directions\nRound 6 text: ask whether to generate the final micro-polish batch now\nRound 6 image action: use OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation for the micro-polish batch\nRound 7: final figure + short legend writing pass\n\n\n## Example 4 — Required navigation footer pattern\n\nEvery text-only planning turn should end with a navigation footer such as:\n\n- **Next step:** The next action is to generate the structural-skeleton candidate board as a separate image batch.\n- **After the images appear, please choose by image** and comment on composition, clutter, and comparison clarity.\n- **Then I will:** update the state and prepare the density / reviewer-bias board.\n\n\n## Host reminder snippet to include before any generation round\n\n- **If you are using ChatGPT web:** I will keep the next step as a separate native image-generation action under Extended Thinking or the strongest available thinking-assisted path, and you do not need to manually switch to Create image first.\n- **If you are using OpenClaw, Codex, Trae, or another IDE/API host:** the next image round must use OpenAI ChatGPT Images 2.0 or a newer supported OpenAI image model. If your host does not have an OpenAI API key configured yet, please add one before we generate the next candidate board.\n\n\n## Example 5 — Required end-of-text reply footer\n\nEvery planning reply should end with something like:\n\n- **Next step:** If you are ready, the next action is to generate the next candidate board as a separate image batch.\n- **After the images appear, please choose by image** and tell me what to keep or change about layout, density, mechanism clarity, reviewer-friendliness, and comparison strength.\n- **Then I will:** update the saved state, record your choice, and prepare the next narrower batch.\n- **For the next turn:** please explicitly ask `paper-framework-figure-studio-pro` to continue from the current saved state when you give your next instruction.\n\n\n## 强制交互节奏（新增）\n\n1. 文字总结当前状态\n2. 文字说明下一批候选图会比较什么\n3. 文字询问：是否现在生成这一批图\n4. 用户确认后，在下一独立生图动作中再调用 OpenAI ChatGPT Images 2.0 / Create image\n5. 生图完成后，再进入下一轮文字评价与选择\n\nFile v1.0.3:assets/figure_brief_template.md\n\n# Figure Brief Template\n\nUse this template immediately after reading the user's paper deep-reading report or model introduction.\n\n## 1. Project identity\n- Working figure title:\n- Paper or method title:\n- One-sentence claim:\n- Source type: deep-reading report / method intro / mixed\n\n## 2. What this figure must explain\n- Core scientific message:\n- What the reviewer should understand after only 10 seconds:\n- What the reviewer should understand after 60 seconds:\n\n## 3. Mandatory content blocks\nList the modules that must appear in the framework figure.\n\n- Problem setting:\n- Inputs / data setting:\n- Main architectural branches:\n- Key learning mechanism:\n- Aggregation or coordination mechanism:\n- Baseline comparison needed?: yes / no\n- Outcomes or benefits box needed?: yes / no\n\n## 4. Boundaries\n- What should stay outside the figure:\n- What can move to caption instead of inside the image:\n- What is too detailed for the main framework figure:\n\n## 5. Audience and venue assumptions\n- Venue style target:\n- Reviewer familiarity: technical / mixed / broader\n- Safety vs novelty preference:\n- Preferred figure density: low / medium / high\n\n## 6. Visual requirements\n- Page format: A4 portrait / A4 landscape / undecided\n- Main color semantics:\n- Must-keep symbols or equations:\n- Mini plots allowed?: yes / no\n- Avatars allowed?: yes / no\n- Style references provided?: yes / no\n\n## 7. Initial recommendation\n- Recommended style family:\n- Recommended backup family:\n- Recommended structural skeleton:\n- Biggest risk to avoid:\n\nFile v1.0.3:assets/next_step_navigation.md\n\n# Next-Step Navigation Patterns\n\nThis file defines how every text-only reply should guide the user into the next one or two steps.\n\n## Principle\n\nA good framework-figure studio conversation never leaves the user guessing. After each text reply, the assistant should explicitly say:\n\n- what will happen next\n- whether the next move is a separate image-generation step\n- what the user should evaluate in the generated candidates\n- what the assistant will do after the user chooses\n\n## Default footer template\n\nUse a concise block like this at the end of every text-planning reply:\n\n**Next step**\n- The next action is: [describe the separate image-generation step or text decision].\n- After the images appear, please choose by image and comment on: [2 to 5 concrete evaluation axes].\n- Then I will: [describe the next narrower step].\n- If you are not sure how to continue after the images appear, you can simply type **\"接下来做什么\"** and I will tell you the recommended next prompt or decision.\n- **Rendering rule reminder:** In ChatGPT web, the next image step must use the assistant's native image generation under the strongest available thinking-assisted path (prefer Extended Thinking when available) and you should **not** manually switch to Create image; in OpenClaw / Codex / Trae / API hosts, the next image step must use **OpenAI ChatGPT Images 2.0** or newer; **SVG and other vector-code fallbacks are forbidden**.\n\n## Good evaluation axes after images appear\n\nDepending on the round, ask the user to comment on some of the following:\n\n- overall style family fit\n- layout / composition\n- visual hierarchy\n- density / clutter\n- reviewer-friendliness\n- mechanism clarity\n- baseline-vs-ours comparison strength\n- icon vocabulary / avatar usage\n- mini scatterplots or result snapshots\n- equation visibility and readability\n- journal-like vs conference-like tone\n\n## Round-specific footer suggestions\n\n### After intake / Figure Brief\n- The next action is to generate a style-family candidate board as a separate image batch.\n- After the images appear, please choose by image and comment on whether you prefer more conservative, modular, mechanism-explanation, flat-illustration, or premium-polish directions.\n- Then I will narrow the winning style into structural skeleton candidates.\n\n### After style-family selection\n- The next action is to generate structural-skeleton candidates as a separate image batch.\n- After the images appear, please choose by image and comment on whether you prefer a left-to-right pipeline, top-down narrative, central-core-with-callouts, or tile-grid composition.\n- Then I will prepare a density / reviewer-bias comparison if needed.\n\n### After structural selection\n- The next action is to generate density / reviewer-bias variants as a separate image batch.\n- After the images appear, please choose by image and comment on technicality, readability, and whether the figure should target expert reviewers or broader readers.\n- Then I will refine the internal visual language.\n\n### After density / bias selection\n- The next action is to generate internal visual-language variants as a separate image batch.\n- After the images appear, please choose by image and comment on avatars, mini-plots, result snapshots, equation count, and comparison-panel strength.\n- Then I will prepare the first integrated exploration batch.\n\n### After exploration batch selection\n- The next action is to generate a narrower refinement batch as a separate image batch.\n- After the images appear, please choose by image and comment on what must stay fixed and what still needs improvement.\n- Then I will lock the final direction and ask whether you want figure text support.\n\n### After final-direction lock\n- The next action is a text-only step: I can draft the caption, legend, and panel explanation text.\n- Please tell me whether you want caption only, legend only, panel callouts, or all three.\n- Then I will produce the publication-facing figure text package.\n\n\n## Resume reminder line\n\nAt the end of every text-only planning reply, add one short continuity reminder such as:\n\n- **For the next turn:** please explicitly ask `paper-framework-figure-studio-pro` to continue from the current saved state when you send the next change request.\n\n## First-contact navigation pattern\n\nUse a first-turn ending such as:\n\n- **Next step:** If you are ready, send your Markdown deep-reading report or your current method description, and I will extract the Figure Brief.\n- **After I read it:** the first image round will usually be a multi-style candidate board for you to choose from.\n- **Then I will:** record the initial state and prepare the first visual decision round.\n- **For the next turn:** please explicitly ask `paper-framework-figure-studio-pro` to continue from the current saved state.\n\n## Mandatory help reminder\n\nEvery text-only planning reply should end with a short help reminder such as:\n\n- **If you are not sure what to ask after the images are generated, just type `接下来做什么`, and I will guide you to the next step.**\n- **If you are unsure how to continue, simply reply `接下来做什么` and I will suggest the next prompt.**\n\n\n## Mandatory per-reply rendering reminder\n\nEvery text-only reply must visibly repeat a short rendering reminder. Do not assume the user still remembers it from earlier turns.\n\nRecommended one-line reminder:\n\n- **Rendering rule reminder:** ChatGPT web should use native image generation under the strongest available thinking-assisted path (prefer Extended Thinking when available) without asking you to manually switch to Create image; OpenClaw / Codex / Trae / API hosts must use OpenAI ChatGPT Images 2.0 or newer; SVG fallbacks are forbidden.\n\nFile v1.0.3:assets/prompt_library.md\n\n# Rendering-path reminder\n\nBefore any generation batch, add a one-line reminder that framework figures must be rendered with OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation, never SVG. In ChatGPT web, keep the user in chat and do not ask them to manually switch to Create image first; in IDE/API hosts, ask for an OpenAI API key if one is missing.\n\n# Prompt Library for Framework Figures\n\nUse these prompts as **building blocks**. Always customize them with the current conversation state.\n\n---\n\n## Prompt Assembly Header\n\nStart prompts with a concrete scientific framing block like this:\n\n> Create a publication-quality A4 [portrait/landscape] framework figure for a top computer science paper. The paper idea is: \"[TITLE]\". The figure must let a reviewer understand the full method at a glance. The figure should be self-contained, highly legible, publication-ready, and suitable for a top-tier conference or journal.\n\nThen add:\n\n- the core claim\n- the required panel sequence\n- the selected style family\n- the visual vocabulary\n- the typography requirements\n- the comparison requirements\n- the output constraints\n\n---\n\n## Candidate-board rule\n\nWhen the user is choosing between schemes, prompts must explicitly create **multiple visually distinct candidates** for the same scientific content.\n\nA good board prompt must state:\n\n- what stays fixed across all candidates\n- what specific axes vary across A / B / C / D\n- that the result should be suitable for user selection by looking at the images\n\nUseful batch instruction pattern:\n\n> Generate [N] candidate framework figures for the same paper. Keep the scientific content fixed. Vary only the following visible axes: [STYLE / LAYOUT / DENSITY / VISUAL LANGUAGE / SNAPSHOT INTENSITY]. Make each candidate clearly distinguishable so the user can choose from the images.\n\n---\n\n## Family 1 — Academic Conservative\n\n> Create a publication-quality A4 [portrait/landscape] scientific framework diagram in a classic top-tier machine learning paper style. Use crisp vector graphics, a white background, light panel borders, restrained color coding, clean arrows, and concise labels. Use blue for shared structure, orange for personal structure, and green for decentralized collaboration. Build a clear [left-to-right / top-down] reading path. Include these panels: (1) problem setting, (2) dual-channel pseudo-labeling, (3) gating, (4) shared-personal bi-timescale learning, (5) selective aggregation, (6) consensus-only vs beyond-consensus comparison, (7) compact outcomes. Add at most one or two equations and keep all text sharp and readable.\n\nBest when:\n- the user wants safety and formal clarity\n- the target audience is technically experienced\n\n---\n\n## Family 2 — Modern Modular Tiles\n\n> Create a publication-quality A4 [portrait/landscape] framework figure in a modern modular magnetic-tile style. Use elegant rounded rectangles, a premium editorial dashboard feel, consistent card spacing, subtle depth, clean typography, capsule labels, and a strong information hierarchy. Organize the figure as modular tiles that a reviewer can scan quickly: title tile, problem-setting tile, shared pseudo-label tile, personal pseudo-label tile, gating tile, shared-branch tile, personal-branch tile, selective-aggregation tile, comparison tile, and outcomes tile. Keep the visual system modern and neat, like a 2025 ML paper figure.\n\nBest when:\n- the user wants modernity without becoming too illustrative\n- the reviewer should be able to scan quickly\n\n---\n\n## Family 3 — Mechanism + Result Snapshots\n\n> Create a publication-quality A4 [portrait/landscape] framework figure in a mechanism-plus-result-snapshot style. Each stage should show both the mechanism and a tiny local effect or state after that stage. Use mini scatterplots, micro heatmaps, tiny decision-boundary sketches, or compact before/after inserts only where they clarify the method. The figure should clearly show: the decentralized problem setup, shared pseudo-label generation, personal pseudo-label generation, sample-wise gating, bi-timescale learning, and selective aggregation. Include concise ribbons such as \"result after this step\" where helpful. Keep the style elegant, not busy.\n\nBest when:\n- the user wants mechanism explainability\n- the method is abstract and benefits from tiny state-change examples\n\n---\n\n## Family 4 — Editorial Flat Illustration\n\n> Create a publication-quality A4 [portrait/landscape] framework figure in a modern editorial flat-illustration style common in recent high-quality ML papers. Use rounded shapes, refined icons, tasteful color blocks, crisp typography, and slight personality without becoming childish. If appropriate, use small client avatars or stylized client icons. Keep all scientific semantics precise. The figure should visually communicate that the method shares what is common while preserving what is personal. Use clear mini plots for local decision boundaries and a polished comparison between consensus-only and beyond-consensus behavior.\n\nBest when:\n- the user wants a memorable main figure\n- the paper also benefits from project-page or talk reuse\n\n---\n\n## Family 5 — Premium Scientific Illustration\n\n> Create a publication-quality A4 [portrait/landscape] framework figure in a high-end scientific illustration style with subtle gradients, soft dimensionality, refined capsules and nodes, thin precise arrows, and sophisticated editorial polish. It should feel premium and memorable while remaining journal-appropriate. Emphasize the shared vs personal structure split, the gating mechanism, the bi-timescale learning, and selective aggregation. Use beautifully rendered network nodes, tidy mini plots, and elegant callout bubbles.\n\nBest when:\n- the user wants a flagship visual\n- the figure may also be used for posters, slides, or project pages\n\n---\n\n## Required content block for this paper direction\n\nFor the current paper idea, the prompt should usually specify these blocks explicitly:\n\n1. Problem setting\n   - decentralized client graph\n   - no central server\n   - few labeled and many unlabeled samples\n   - heterogeneous client-specific structures\n   - note that not all differences are noise\n\n2. Dual-channel pseudo-labeling\n   - shared pseudo-label from neighbors\n   - personal pseudo-label from local model\n\n3. Sample-wise gate\n   - show \\( g_i(x) \\in [0,1] \\)\n   - include the fusion equation\n\n4. Shared-personal bi-timescale learning\n   - shared branch \\(\\theta_i^{sh}\\)\n   - personal branch \\(\\theta_i^{per}\\)\n   - fast collaborative update vs slow local adaptation\n\n5. Selective aggregation\n   - only shared branch aggregated\n   - weight logic includes quality, similarity, personalization risk\n\n6. Beyond-consensus comparison\n   - consensus-only overwrites personal structure\n   - our method preserves client-specific structure while still collaborating\n\n7. Outcomes\n   - better average accuracy\n   - stronger worst-client performance\n   - higher personalization retention\n\n---\n\n## Standard equation snippets\n\nUse only a few equations inside the figure. Good candidates:\n\n- \\( \\hat{y}_i(x) = g_i(x)\\hat{y}_i^{sh}(x) + (1-g_i(x))\\hat{y}_i^{per}(x) \\)\n- \\( \theta_i^{sh,t+1} = \\sum_j w_{ij}^{sh,t}\\,\tilde{\theta}_j^{sh,t} \\)\n- weight hint: quality + similarity - personalization risk\n\n---\n\n## Style-family board prompt pattern\n\n> Create [N] distinct candidate framework figures for the same paper idea. Keep the scientific modules fixed. Candidate A should emphasize [AXIS 1]. Candidate B should emphasize [AXIS 2]. Candidate C should emphasize [AXIS 3]. Candidate D should emphasize [AXIS 4]. Make the candidates visually different enough for image-based selection. Do not change the scientific claim.\n\nExample varying axes:\n- classic conservative vs modular tiles vs mechanism snapshots vs flat illustration\n- more formal vs more modern vs more explanation-heavy vs more visually memorable\n\n---\n\n## Structural-skeleton board prompt pattern\n\n> Create [N] candidate framework figures for the same paper and same style family, but vary only the composition skeleton. Candidate A: left-to-right pipeline. Candidate B: top-down narrative stack. Candidate C: central mechanism with surrounding callouts. Candidate D: modular tile grid. Keep typography family, color semantics, and scientific content fixed. Make layout differences obvious enough for image-based selection.\n\n---\n\n## Density / reviewer-bias board prompt pattern\n\n> Create [N] candidate framework figures with identical scientific content and same overall style direction, but vary the visual density and reviewer orientation. Candidate A: technical/formal medium density. Candidate B: cross-domain easier-to-understand medium density. Candidate C: visually modern but still rigorous medium density. Candidate D: high-density expert version. Keep the content fixed and make the density differences obvious in the images.\n\n---\n\n## Internal visual-language board prompt pattern\n\n> Create [N] candidate framework figures that keep the chosen layout stable but vary the internal visual language. Candidate A: abstract nodes, no avatars, minimal mini-plots. Candidate B: small avatars or client icons with mini scatterplots. Candidate C: result snapshots and richer callouts. Candidate D: fewer labels but one stronger equation panel. Keep the paper content fixed and make the differences visible for image-based selection.\n\n---\n\n## First integrated exploration-batch prompt pattern\n\n> Create [N] candidate framework figures for the same paper idea. Keep the scientific content fixed. Vary only the selected exploration axes from the current state: [AXES]. All candidates must be publication-ready, A4-balanced, and legible. The candidates must be different enough that a user can choose among them by looking at the images.\n\n---\n\n## Refinement-batch prompt pattern\n\n> Refine the selected direction while keeping its core composition. Preserve [KEEP LIST]. Change only the following aspects: [CHANGE LIST]. Improve label hierarchy, reduce clutter, and make the main claim more immediately legible. Keep the same paper content and A4 publication balance. Produce [N] refinement variants that differ only in controlled visible ways so the user can choose from the images.\n\n---\n\n## Final-polish prompt pattern\n\n> Create a final polished framework figure based on the chosen direction. Keep the composition stable. Tighten spacing, sharpen text, unify icon style, improve panel hierarchy, and ensure the figure looks ready for a top conference or journal submission. Do not introduce new modules. Make the final comparison panel crisp and reviewer-friendly.\n\n\n## Text-only reply footer rule\n\nAfter every text-only planning reply, add a short navigation footer that tells the user:\n\n- what the next action is\n- whether the next action is a separate image-generation step\n- what feedback to give after the images appear\n- what the assistant will do after the user chooses\n\nSuggested footer pattern:\n\n> **Next step:** The next action is to generate [BOARD TYPE] as a separate image batch.\n> **After the images appear, please choose by image** and comment on: [AXIS 1], [AXIS 2], [AXIS 3].\n> **Then I will:** [NEXT NARROWER STEP].\n\n\n## First-contact text scaffold\n\nUse this on the first planning turn before any image generation:\n\n1. Recommend a Markdown deep-reading report as the preferred input.\n2. Mention the paper-deep-reading skill URL: `https://clawhub.ai/c-narcissus/paper-deep-reading`.\n3. Clarify that this is recommended, not required.\n4. Say that method sketches, module descriptions, algorithm notes, and early design ideas are also acceptable.\n5. Ask whether the user is ready to start figure design now.\n6. Preview that after reading the input, the first generation round will usually be a multi-style candidate board.\n7. End with a Next Steps block and a resume reminder.\n\nFile v1.0.3:assets/refinement_controls.md\n\n# Refinement Controls\n\nUse these controls after a user selects a winner or shortlist.\n\n## Hierarchy\n- make the main claim more dominant\n- reduce secondary panels by one visual level\n- increase white space between conceptual groups\n- enlarge panel titles and compress body labels\n\n## Clarity\n- simplify arrows\n- reduce repeated icons\n- merge two small panels into one clearer panel\n- move low-value labels out of the image and into the caption\n\n## Technicality\n- add one clean formula snippet\n- remove formulas and rely on visual logic only\n- make the comparison panel more formal\n- make the baseline-vs-ours difference more explicit\n\n## Style tuning\n- less playful, more journal-like\n- more modern, less conservative\n- flatter and cleaner\n- more premium scientific illustration\n- fewer avatars, more abstract nodes\n- more modular tile feel\n\n## Density\n- compress into fewer panels\n- expand one key stage with more detail\n- add mini result snapshots\n- remove mini result snapshots\n\n## Layout\n- switch to A4 portrait balance\n- switch to A4 landscape balance\n- make the top area lighter and comparison area tighter\n- enlarge the central mechanism and shrink the legend area\n\n## Choice protocol reminder\n\nWhen any of the above controls would produce a visibly different figure direction:\n\n1. summarize the proposed refinement axis in text\n2. ask whether to generate the next refinement candidate board now\n3. generate 2 to 4 visible variants in a separate image action\n4. ask the user to choose from the resulting images\n\nArchive v1.0.2: 21 files, 38480 bytes\n\nFiles: assets/conversation_state.template.json (4582b), assets/figure_brief_template.md (1527b), assets/next_step_navigation.md (5717b), assets/prompt_library.md (11863b), assets/refinement_controls.md (1517b), CHANGELOG.md (2614b), examples/example_opening_turns.md (2816b), publish/cover_and_icon_prompts.md (817b), publish/listing_long.md (1294b), publish/listing_short.md (317b), publish/release_checklist.md (555b), publish/starter_messages.md (1580b), README.md (5234b), references/clawhub_packaging_notes.md (1413b), references/README_CN.md (7904b), references/reviewer_style_taxonomy.md (871b), references/visual_communication_principles.md (1765b), references/workflow_examples.md (8092b), SKILL.md (27077b), templates/user_input_bundle.md (863b), _meta.json (152b)\n\nFile v1.0.2:SKILL.md\n\n---\nname: paper-framework-figure-studio-pro\ndescription: Convert a paper deep-reading report, method description, or model introduction into publication-ready framework figure concepts through a stateful multi-round workflow. Use when the user wants top-conference/top-journal framework diagrams, style exploration, human approval between rounds, concrete image prompts, separate text-versus-image turns, and iterative refinement toward a final paper figure.\ncompatibility: ChatGPT web, Codex, Trae, OpenClaw, ClawHub marketplace, skills.sh. Tool-agnostic. Requires an image-generation capability in the host environment for rendering.\nmetadata:\n  display_name: Paper Framework Figure Studio Pro\n  version: \"1.4.3\"\n  author: OpenAI\n  tags: research-figure, scientific-illustration, framework-diagram, paper-writing, clawhub\n---\n\n# Paper Framework Figure Studio Pro\n\nTurn a paper's deep-reading report, model introduction, or method summary into a publication-ready **framework figure** through a **stateful, multi-round, human-in-the-loop studio workflow**.\n\nUse this skill when the user wants any of the following:\n\n- a main paper framework diagram\n- a method overview figure for a conference or journal paper\n- multi-style exploration of academic figure directions\n- iterative narrowing from broad style families to polished final renders\n- concrete, detailed prompts for image generation rather than vague design suggestions\n- explicit human confirmation between rounds\n- a final optional pass that writes the figure legend / caption / panel callouts\n\nThis skill is primarily for **framework figures**, not benchmark plots, tables, or raw quantitative charts.\n\n## Core operating model\n\nThis skill is not a one-shot prompt generator. It is a **conversation-driven figure studio** with explicit state.\n\n### Mandatory turn separation protocol\n\nEvery generation cycle must follow this order:\n\n1. **Text-only planning / summary turn**\n2. **Text-only confirmation turn** asking whether to generate the next candidate batch now\n3. **Image-only generation action/turn** after the user says yes\n4. **Text-only evaluation turn** asking the user to choose from the generated images\n\nIf the host environment tends to auto-generate images together with text, the skill must still behave as if those are forbidden to co-occur, and should explicitly defer image generation to the next turn after confirmation.\n\n## First-contact protocol\n\nOn the **first planning turn**, before the studio fully starts, the assistant should briefly orient the user.\n\n### Preferred but not mandatory upstream input\n\nThe assistant should explicitly tell the user that the best upstream input is usually a paper deep-reading report in Markdown, and should recommend the user first use the **paper-deep-reading** skill for the target paper or draft when available.\n\nRecommended reminder wording should communicate all of the following:\n\n- best practice is to first generate a paper deep-reading report or structured reading report\n- the report should preferably be saved in **Markdown**\n- the user may use the paper-deep-reading skill at `https://clawhub.ai/c-narcissus/paper-deep-reading` if they want a strong upstream report\n- this is **recommended, not required**\n- the skill also accepts less-complete inputs such as a method sketch, module list, algorithm description, or early-stage design notes\n\n### First-turn readiness check\n\nAfter the recommendation, the assistant should explicitly ask whether the user is **ready to start figure design now**.\n\nThe first-turn planning reply should therefore do four things in order:\n\n1. remind the user that a Markdown deep-reading report is the preferred input, though not mandatory\n2. state what kinds of partial inputs are also acceptable\n3. ask whether the user is ready to begin figure design now\n4. preview that once the report or description is read, the first image round will usually be a **multi-style candidate board** for the user to choose from\n\n### First turn after ingesting the report or method description\n\nOnce the assistant has read the user's deep-reading report or method description and extracted the Figure Brief, it should explicitly tell the user that the normal next move is to generate **one batch of multiple style directions** for visual comparison.\n\nThat first post-ingest planning turn should:\n\n- summarize the extracted Figure Brief\n- name the first decision as a **style-family decision**\n- explain that the decision should be made by looking at generated candidate images rather than prose only\n- ask whether to generate the first multi-style candidate board now\n\n\nThe assistant should:\n\n1. read the user's paper deep-reading report, method description, or model summary\n2. extract a clean **Figure Brief**\n3. open a multi-round workflow\n4. after each round, summarize the current state and ask for one concrete user decision\n5. **separate all text planning from all image generation**\n6. use **image candidates as the actual decision surface** whenever the user is choosing between visual schemes\n7. update the running state after every user choice and every generation batch\n8. progressively narrow style, structure, density, and detail\n9. after the user selects a final direction, ask whether to also draft the **figure caption / legend / panel explanation text**\n10. end every text-planning reply with a short **Next Steps** block so the user always knows what the next one or two actions will be\n11. after every text-planning reply, record the updated session state, including generated deliverables and user selections\n12. remind the user that future turns should explicitly ask to continue with this skill based on the current saved state\n13. remind the user that if they are unsure what to ask next after images are generated, they can simply type **\"接下来做什么\"** (or **\"what should we do next\"**) to receive guided next-step instructions\n\n## Non-negotiable rules\n\n- **Human approval gate required** between major rounds.\n- **Do not jump straight to a final image** from the initial paper description.\n- **Do not silently change the chosen direction** without telling the user.\n- **Do not mix planning text and image generation in the same reply** when the host supports separate image actions. First do the text reply. Then do image generation as a distinct action.\n- **Treat this as a hard runtime constraint:** if a reply contains explanation, summary, questions, next-step guidance, or confirmation requests, that reply must be **text-only** and must not trigger image generation.\n- **Before every image batch there must be a dedicated confirmation reply** whose sole job is to ask whether the user wants to generate that batch now. The actual image-generation action must happen only after that confirmation, as a separate next action/turn.\n- **Do not use SVG as the primary or fallback rendering path** for framework-figure candidate boards or finals.\n- **For any round where the user must choose among visual schemes, do not ask them to choose only from prose descriptions. Generate visual candidates first, then ask them to choose by looking at the images.**\n- **Before each new generation batch, explicitly ask whether the user wants to generate the next set of candidate images now.**\n- **Before each generation batch, explicitly name the rendering path that will be used in the current host: ChatGPT web should use native Create image via the assistant under Extended Thinking or the strongest available thinking-assisted path; IDE/API hosts should use OpenAI ChatGPT Images 2.0 or a newer supported OpenAI image model.**\n- **Always track state** using a structure equivalent to `assets/conversation_state.template.json`.\n- **On the first turn, recommend but do not require a Markdown deep-reading report created with the paper-deep-reading skill before figure work begins.**\n- **After reading the report or method description for the first time, explicitly propose generating a first multi-style candidate board for selection.**\n- **After every text-only reply, update the running state with the current figure brief, generated deliverables, pending decisions, and the user's recorded preferences.**\n- **At the end of every text-only reply, remind the user that later messages should explicitly ask this skill to continue from the current state.**\n- **At the end of every text-only planning reply, explicitly remind the user that if they do not know how to continue after the next image batch, they can simply type `接下来做什么` to get guided prompt suggestions and next-step help.**\n- **At the end of every text-only reply, explicitly restate the rendering rule for the current host:** ChatGPT web should use the assistant's native image generation under the strongest available thinking-assisted path (prefer Extended Thinking when available) without asking the user to manually switch to Create image; IDE/API hosts must use OpenAI ChatGPT Images 2.0 or a newer supported OpenAI image model; SVG and other vector-code fallbacks are forbidden.\n- **Generate multiple candidates per batch**. Early exploration batches should usually produce 3 to 5 candidates. Later refinement batches should usually produce 2 to 4 candidates.\n- **Make prompts concrete**. Avoid vague instructions like “make it look academic” unless followed by precise layout, hierarchy, color, metaphor, and typography requirements.\n- **Prefer framework-figure clarity over decorative complexity**.\n- **Ask whether the user wants figure text help after the final image direction is chosen**.\n- **At the end of every text-only planning reply, explicitly tell the user what the next one or two steps are, whether the next step is a text decision or an image-generation step, and what kinds of feedback they should be ready to give after the images appear.**\n\n## Host-specific image-generation policy\n\nWhen this skill reaches a generation step, use the host environment's best available **native OpenAI image-generation path**, and keep it separate from the text-planning reply.\n\nThe separation rule is absolute: a planning / explanation / confirmation message and an image-generation action must never be bundled into one assistant reply.\n\n### Hard prohibition\n\n- **Do not use SVG as the rendering path for candidate boards or final framework figures.**\n- **Do not switch to code-drawn vector output as a substitute for native image generation.**\n- **Do not use mermaid, tikz, graphviz, or other vector-code fallbacks for framework-figure rendering rounds.**\n- The intended rendering path is **OpenAI Create image / ChatGPT Images**, specifically **ChatGPT Images 2.0 or a newer supported OpenAI image model if the host exposes one**.\n\n### ChatGPT web\n\n- The preferred interaction is: the user stays inside the normal chat and the assistant triggers the host's native image generation as a separate image action.\n- **Do not ask the user to manually switch tools or manually click a Create image mode first**; the skill should treat image generation as a native follow-up action after the planning reply.\n- Prefer **Extended Thinking** for framework-figure generation when the host exposes it.\n- If the host experience exposes **Thinking** or **images with thinking** but not an explicit Extended Thinking label, prefer the strongest available reasoning-assisted image path.\n- Treat the image step as its own action after the user says to proceed.\n\n### OpenClaw, Codex, Trae, or other IDE / API-driven hosts\n\n- Use **OpenAI ChatGPT Images 2.0** at minimum for raster image output.\n- If the host exposes a newer OpenAI image-generation version than ChatGPT Images 2.0, use the newer supported OpenAI version.\n- If the host requires an API key and no OpenAI API key is available, **pause before generation and explicitly tell the user that image generation cannot proceed until they provide or configure an OpenAI API key**.\n- When generation is blocked by missing credentials, do not fake progress and do not switch to SVG as a fallback.\n- Do **not** replace image generation with SVG, mermaid, tikz, graphviz, or other vector-code fallbacks when the task is a framework-figure rendering step.\n\n### Required interaction split\n\n- Keep the **text planning reply** and the **image-generation action** separate.\n- When a round is a **visual decision round**, the normal sequence is:\n  1. text turn: summarize state, remind the user of the rendering path that will be used in this host, and ask whether to generate the next candidate board\n  2. image action: use **OpenAI Create image / ChatGPT Images 2.0 or newer supported OpenAI image generation** to generate the candidate board or multi-image batch\n  3. text turn: briefly label the shown candidates and ask the user to choose by image number / letter\n\n## Visual-decision-first protocol\n\nThis skill uses a **visual-decision-first** workflow.\n\nThat means:\n\n- The assistant may explain what varies across options in text.\n- But when the user's next decision is fundamentally about figure appearance, layout, style family, internal visual language, or refinement direction, the assistant should not stop at text-only options.\n- Instead, the assistant should ask whether to generate a **candidate board** for that decision.\n- After generation, the assistant should present a short mapping such as **A / B / C / D** or **1 / 2 / 3 / 4** tied to the generated images and ask the user to choose from the images.\n\nGood visual-decision rounds include:\n\n- style family selection\n- structural skeleton selection\n- density / audience-bias tradeoff when it changes the figure look\n- internal visual language selection\n- refinement direction selection\n- final shortlist selection\n\n\n## Mandatory next-step navigation\n\nEvery **text-only** reply in the workflow must end with a short navigation block.\n\nThe navigation block should be concrete and user-facing, not abstract process language. It should tell the user:\n\n1. what the very next step is\n2. whether that next step is **another text decision** or a **separate image-generation action**\n3. what the user will need to do right after images are generated\n4. what kinds of feedback will be most useful in the next turn\n\nIt should also repeat one short **rendering-rule reminder** in every text-only reply so the user sees it every round:\n\n- In **ChatGPT web**, image generation must be a separate native image action under the strongest available thinking-assisted path; prefer **Extended Thinking** when available, and do **not** ask the user to manually switch to Create image.\n- In **OpenClaw / Codex / Trae / API hosts**, image generation must use **OpenAI ChatGPT Images 2.0** at minimum, or a newer supported OpenAI image model if available.\n- **SVG, mermaid, tikz, graphviz, and other vector-code fallbacks are forbidden.**\n\n### Required structure\n\nUse a compact structure such as:\n\nIn addition to the user-facing navigation block, the assistant should also internally update the saved session state after the text reply is composed. That state update should include:\n\n- the latest accepted inputs\n- the current Figure Brief\n- every generated deliverable so far\n- the most recent user choices and rejected options\n- the currently pending decision\n- the next candidate board that would be generated if the user says yes\n\n\n- **Next step:** [ask permission to generate the next candidate board / summarize a shortlisted direction / write caption text]\n- **After the images appear, please choose by image** and optionally comment on: [layout / density / comparison clarity / icon style / equations / mini-result snapshots / clutter / reviewer-friendliness]\n- **Then I will:** [update the chosen direction and prepare the next narrower batch]\n- **For the next turn:** please explicitly ask `paper-framework-figure-studio-pro` to continue from the current saved state.\n\n### Session continuity reminder\n\nBecause some hosts do not automatically preserve skill-specific working memory in a reliable way, the assistant should remind the user at the end of each text-only planning turn that future messages should explicitly say something like:\n\n- “Please continue with **Paper Framework Figure Studio Pro** from the current saved state.”\n- “Use **paper-framework-figure-studio-pro** to continue from the current state and apply this new change request.”\n\nThis reminder should be brief, but it should appear consistently so the user knows how to resume the workflow in later turns. The assistant should also record the updated session state after every text-planning reply, including generated deliverables and user selections.\n\n### Examples\n\nExample A:\n\n- **Next step:** If you want, the next action is to generate the style-family candidate board as a separate image batch.\n- **After the images appear, please choose by image** (A/B/C/D) and tell me what you like or dislike about layout, modernity, and explanation strength.\n- **Then I will:** update the state and prepare the next structural-skeleton board.\n\nExample B:\n\n- **Next step:** The next action is to generate a refinement batch focused only on reducing clutter and strengthening the baseline-vs-ours comparison.\n- **After the images appear, please choose by image** and note whether you want fewer labels, cleaner arrows, or stronger mini-result snapshots.\n- **Then I will:** lock the winning direction and ask whether you also want caption / legend / panel text.\n\nDo not omit this navigation block. The user should always know the next one or two moves.\n\n## Workflow overview\n\nFollow this sequence unless the user explicitly asks to skip or compress a stage.\n\n### Round 0 — Intake and figure brief construction\n\nRead the user's deep-reading report or model description and construct a **Figure Brief**.\n\nThe Figure Brief must capture at least:\n\n- paper or method title\n- one-sentence scientific claim\n- what the figure must explain\n- target figure type: framework overview\n- likely venue level and audience familiarity\n- mandatory modules to show\n- optional modules to compare\n- what should remain outside the figure\n- preferred page format: A4 portrait, A4 landscape, or unknown\n- whether the user values safety, modernity, mechanism explanation, or visual memorability more\n\nUse the template in `assets/figure_brief_template.md`.\n\n### Round 1 — Style-family candidate board\n\nDo **not** ask the user to choose only from a written list of families.\n\nInstead:\n\n1. summarize 3 to 5 candidate families very briefly in text\n2. ask whether to generate the **style-family candidate board now**\n3. in a separate image action, generate 3 to 5 style-distinct figure candidates for the same paper content\n4. after the images are shown, label them in a short text turn and ask the user to choose a primary direction and optionally a backup\n\nRecommended default families:\n\n1. **Academic Conservative** — standard top-tier ML paper overview\n2. **Modern Modular Tiles** — magnetic-card / dashboard-like figure blocks\n3. **Mechanism + Result Snapshots** — each stage shows both mechanism and local effect\n4. **Editorial Flat Illustration** — modern flat/cartoon academic style, friendly but rigorous\n5. **Premium Scientific Illustration** — soft-3D / high-polish scientific editorial rendering\n\n### Round 2 — Structural-skeleton candidate board\n\nWithin the selected family, do not stop at textual skeleton descriptions.\n\nInstead:\n\n1. propose 2 to 4 structural skeletons very briefly\n2. ask whether to generate the **structural-skeleton candidate board now**\n3. in a separate image action, generate 2 to 4 candidates where the content stays fixed but the composition changes, such as:\n   - left-to-right pipeline\n   - top-down narrative stack\n   - central model + surrounding callouts\n   - modular tile grid\n   - comparison split with baseline vs ours\n4. after the images are shown, ask the user to choose from the images\n\n### Round 3 — Density / reviewer-bias candidate board\n\nIf density or reviewer bias will materially affect the visual appearance, do not ask the user to decide only from prose.\n\nInstead:\n\n1. explain that the next board will compare, for example:\n   - technical / formal\n   - cross-domain / easier to understand\n   - visually modern but still rigorous\n   and/or\n   - low density\n   - medium density\n   - high density\n2. ask whether to generate the **density-and-bias candidate board now**\n3. generate 2 to 4 candidates in a separate image action\n4. ask the user to choose from the images\n\n### Round 4 — Internal visual-language candidate board\n\nNarrow the figure's visual language.\n\nTypical selectable elements:\n\n- avatars or no avatars\n- mini scatterplots or no mini scatterplots\n- per-step result snapshots or mechanism only\n- one equation or several small equations\n- minimal labels or richer callout labels\n- baseline comparison included or deferred\n\nProtocol:\n\n1. summarize what will vary\n2. ask whether to generate the **internal-visual-language board now**\n3. generate 2 to 4 candidates in a separate image action\n4. ask the user to choose from the images\n\n### Round 5 — Exploration batch generation\n\nPrepare a concrete batch with 3 to 5 candidate prompts.\n\nImportant:\n\n- keep the paper content fixed\n- vary only a few style axes per batch\n- state clearly what differs across candidates\n- ask whether to generate this batch now\n- after the user approves, perform image generation in a separate action\n- after the images appear, ask the user to choose from the actual images rather than from abstract prose\n\nThen update state with the generated batch metadata.\n\n### Round 6 — Selection and refinement\n\nAfter the user chooses a winner or shortlist:\n\n- summarize what won\n- summarize what the user disliked\n- propose the next refinement axis\n- ask whether to generate the narrower refinement batch now\n- generate the refinement batch in a separate image action\n- after the images appear, ask the user to choose from the actual images\n\nTypical refinement axes:\n\n- stronger hierarchy\n- less clutter\n- more legible equations\n- better baseline-vs-ours comparison\n- cleaner client graph\n- more journal-like typography\n- more modern or less playful icons\n- closer to A4 publication balance\n\n### Round 7 — Finalization\n\nOnce the user selects a final direction:\n\n- confirm the final figure intent\n- ask whether they also want:\n  - panel labels\n  - legend text\n  - figure caption\n  - figure explanation for the paper body\n  - bilingual callout wording\n\n## Mandatory text-turn protocol\n\nEvery non-image reply should follow this pattern.\n\n### A. Current state\n\nBriefly state:\n\n- current round\n- current chosen family and skeleton\n- current unresolved decision\n\n### B. Visual decision to be made\n\nIf the next decision is visual, say that the next step should be based on **candidate images**, not only verbal descriptions.\n\n### C. Ask permission for the next image batch\n\nAsk a bounded confirmation such as:\n\n- “Do you want me to generate the next style-family candidate board now?”\n- “Do you want me to generate the structural-layout candidates now?”\n- “Do you want me to generate the next refinement batch now?”\n\n### D. What the next batch will vary\n\nState 2 to 5 controlled axes that will differ across the generated images.\n\n### E. After images are shown\n\nIn the next text turn after generation:\n\n- label the shown candidates clearly\n- give a one-line difference summary for each candidate\n- ask the user to choose by image ID, for example **A**, **B**, **C**, **D**\n\n## Mandatory image-turn protocol\n\nEvery image-generation step must be independent from the planning text turn.\n\n- Do not include a long discussion inside the image-generation step.\n- The generation action should use a concrete prompt assembled from the current state.\n- Early rounds should produce multiple style-diverse candidates.\n- Later rounds should produce tightly controlled refinements.\n- The batch should be assembled so that the user can make a real choice **from the generated images**.\n\n## Prompt-construction standard\n\nBuild prompts from the following layers, in this order.\n\n1. **Figure goal** — what the figure explains scientifically\n2. **Paper framing** — title and one-line claim\n3. **Required content blocks** — the exact modules that must appear\n4. **Narrative order** — the intended reading path\n5. **Style family** — one of the chosen families\n6. **Structural skeleton** — layout archetype\n7. **Visual vocabulary** — icons, nodes, mini plots, avatars, tiles, cards\n8. **Typography requirements** — concise labels, sharp text, panel headings\n9. **Color semantics** — blue shared, orange personal, green collaboration by default\n10. **Comparative emphasis** — consensus-only vs beyond consensus if included\n11. **Output constraints** — A4, portrait/landscape, publication-ready, uncluttered, legible\n12. **Batch-difference instruction** — what should differ across candidate A/B/C/D and what must stay fixed\n\nUse the detailed templates in `assets/prompt_library.md`.\n\n## Visual-communication standards\n\nFramework figures should satisfy the following principles.\n\n- one dominant message per figure\n- strong reading path\n- consistent visual metaphor\n- stable color semantics across rounds\n- enough white space to separate reasoning chunks\n- panel labels must reflect conceptual boundaries, not arbitrary boxes\n- if mini result snapshots are used, they must illustrate a real conceptual change rather than act as decoration\n- decorative flair must never obscure the core method\n- image batches should differ along deliberate axes that are visible enough for the user to judge from the images\n\n---\nSee `references/visual_communication_principles.md`.\n\n## Common failure modes to avoid\n\n- asking the user to choose a visual direction from text only when images are required to judge it\n- forgetting to ask whether to generate the next candidate board now\n- mixing the explanation turn and the image turn into a single blended reply\n- too much tiny text inside the image\n- mixing too many styles in one batch\n- icons that imply the wrong algorithmic semantics\n- confusing “shared” with “global final model” when the paper is personalized\n- making every step equally visually heavy\n- comparison panel larger than the main mechanism\n- decorative 3D effects that damage legibility\n- using unrealistic benchmark plots when the figure is supposed to be a framework diagram\n- asking the user too many open-ended questions at once\n\n## What to ask after a final figure is chosen\n\nAlways ask:\n\n- Do you want me to also write the **figure caption**?\n- Do you want **panel-wise explanatory text** for the paper body or appendix?\n- Do you want a **short legend / callout wording pass** to improve what appears inside the figure?\n\n## Files in this skill bundle\n\n- `assets/figure_brief_template.md`\n- `assets/conversation_state.template.json`\n- `assets/prompt_library.md`\n- `assets/refinement_controls.md`\n- `references/visual_communication_principles.md`\n- `references/reviewer_style_taxonomy.md`\n- `references/workflow_examples.md`\n- `references/README_CN.md`\n\nUse them actively rather than improvising from scratch every round.\n\nFile v1.0.2:README.md\n\n# Paper Framework Figure Studio Pro\n\n**Paper Framework Figure Studio Pro** is a stateful, multi-round scientific-figure skill for turning a paper deep-reading report, method summary, module sketch, or algorithm description into a publication-ready **framework figure workflow**.\n\nIt is designed for **top-tier CS paper figures** where users want:\n\n- explicit human confirmation between rounds\n- visual candidate boards before style/layout decisions\n- text planning and image generation kept separate\n- OpenAI native image generation only for figure renders\n- recorded state across rounds\n- final optional help with caption, legend, and panel explanation text\n\nThis package is prepared as a **publish-ready skill bundle** for OpenClaw / ClawHub style runtimes and similar hosts.\n\n## What this skill does\n\nThe skill reads the user's paper or method description, builds a **Figure Brief**, then runs a guided studio workflow:\n\n1. Recommend a Markdown deep-reading report as the best upstream input (but do not require it)\n2. Confirm the user is ready to begin figure design\n3. Extract the Figure Brief\n4. Propose the first **multi-style candidate board**\n5. Generate multiple image candidates as a **separate image action**\n6. Ask the user to choose by looking at the generated images\n7. Update state, narrow direction, and continue to the next refinement round\n8. Ask at the end whether the user also wants caption / legend / panel explanation support\n\n## Hard rules\n\n- **No SVG rendering path** for candidate boards or final framework figures\n- **No mermaid / graphviz / tikz fallback** for figure-rendering rounds\n- Text planning and image generation must be **separate steps**\n- **A reply may never contain both planning text and image generation.** If the assistant is asking, explaining, summarizing, or requesting confirmation, that reply must be text-only.\n- **Before each generation batch, there must be a dedicated text-only confirmation turn** asking whether to generate the next candidate images now. The actual image generation must happen only in the next separate action/turn after the user confirms.\n- Visual decisions should be made from **generated images**, not prose-only descriptions\n- Every text turn must end with **Next Steps** guidance\n- Every text turn must update and preserve session state\n\n## Image-generation policy\n\n### ChatGPT web\nUse the host's native **Create image** path as a separate action after the planning reply. Prefer **Extended Thinking** or the strongest available thinking-assisted image path exposed by the host. Do not instruct the user to manually switch tools first.\n\n### OpenClaw / Codex / Trae / IDE / API hosts\nUse **OpenAI ChatGPT Images 2.0** at minimum, or a newer supported OpenAI image model if exposed by the host. If no OpenAI API key is available, the skill must pause and ask the user to provide or configure one before generation.\n\n## Package contents\n\n- `SKILL.md` — main skill specification\n- `LICENSE` — MIT-0 / MIT No Attribution license text\n- `VERSION` — current package version\n- `CHANGELOG.md` — release notes\n- `assets/` — working templates and navigation / prompt libraries\n- `references/` — Chinese guide, workflow notes, reviewer taxonomy, visual communication principles\n- `examples/` — suggested opening turns and continuation patterns\n- `publish/` — release-page copy, listing text, icon / cover prompts, publishing checklist\n- `templates/` — optional user-facing input templates\n\n## Suggested release metadata\n\n- **Slug:** `paper-framework-figure-studio-pro`\n- **Name:** `Paper Framework Figure Studio Pro`\n- **License:** MIT-0\n\n## Quick start\n\nThe best first user message is something like:\n\n> Please use **paper-framework-figure-studio-pro**. I have a deep-reading report in Markdown for my paper draft. Read it, extract the figure brief, and tell me whether we should generate the first multi-style candidate board.\n\nOr, for an early-stage project:\n\n> Please use **paper-framework-figure-studio-pro**. I do not have a full draft yet. I only have a model description and module design notes. Read them, build the figure brief, and tell me whether I am ready to start the first candidate board.\n\n## Recommended upstream companion skill\n\nThis skill works best when the user first prepares a paper reading report in Markdown with **paper-deep-reading**. The skill should recommend, but not require, the upstream deep-reading workflow.\n\n## Release note\n\nThis package is intentionally focused on **framework figures**. It is not a plot generator, chart generator, or general slide-design skill.\n\n- At the end of every text-only planning reply, the studio should remind the user that if they are unsure how to continue after the next image batch, they can simply type **`接下来做什么`** to receive guided next-step instructions.\n\n\nRendering rule reminder used in every text-only reply:\n- ChatGPT web: use native image generation under the strongest available thinking-assisted path; prefer Extended Thinking when available; do not ask the user to manually switch to Create image.\n- OpenClaw / Codex / Trae / API hosts: use OpenAI ChatGPT Images 2.0 or newer.\n- SVG, mermaid, tikz, graphviz, and other vector-code fallbacks are forbidden.\n\nFile v1.0.2:_meta.json\n\n{\n  \"ownerId\": \"kn7fxns1xpr6z67w885my7d7k98506vv\",\n  \"slug\": \"paper-framework-figure-studio-pro\",\n  \"version\": \"1.0.2\",\n  \"publishedAt\": 1776944342489\n}\n\nFile v1.0.2:references/clawhub_packaging_notes.md\n\n# ClawHub Packaging Notes\n\nThis bundle is prepared for ClawHub / OpenClaw style packaging:\n\n- directory name matches the publish slug: `paper-framework-figure-studio-pro`\n- `SKILL.md` uses YAML frontmatter\n- the frontmatter `name` is the lowercase publish-safe skill identifier\n- the human-readable display name is stored in `metadata.display_name`\n- no conflicting license terms are added inside `SKILL.md`\n\nIf a publish UI asks separately for display name, use:\n\n- **Slug**: `paper-framework-figure-studio-pro`\n- **Name / Display Name**: `Paper Framework Figure Studio Pro`\n\n\n## Image-rendering constraint\n\nThis skill is intentionally authored for **native image generation** rather than SVG synthesis. In hosts that support separate image actions, the recommended path is a distinct **Create image** step, ideally using **Thinking** or **Extended Thinking / images with thinking** when available for complex framework figures.\n\n\nAdditional host policy for v1.2.1:\n- ChatGPT web: keep the user in chat, prefer Extended Thinking or the strongest available thinking-assisted image path, and do not ask the user to manually switch tools before generation.\n- IDE / API hosts such as OpenClaw, Codex, and Trae: use OpenAI ChatGPT Images 2.0 at minimum; if credentials are missing, stop and ask the user to configure an OpenAI API key before any generation round.\n- Never downgrade framework-figure rendering to SVG.\n\nFile v1.0.2:references/README_CN.md\n\n# 发布版说明\n\n当前目录已经补齐为一个更完整的发布版 skill 包，额外包含：\n\n- `README.md`：英文发布页说明\n- `LICENSE`：MIT-0 正式授权文本\n- `CHANGELOG.md`：版本记录\n- `examples/`：示例开场与续接写法\n- `publish/`：ClawHub / OpenClaw 上架页文案、封面与图标 prompt、发布检查清单\n- `templates/`：用户输入模板\n\n---\n\n# Paper Framework Figure Studio Pro 使用说明（中文）\n\n这是一个面向论文**框架图**的多轮科研绘图 skill。\n\n它不是“直接给一条提示词然后开画”，而是：\n\n1. 先读取用户的论文精读报告、方法介绍或模型说明\n2. 抽取 Figure Brief\n3. 开启多轮对话\n4. 每一轮只让用户做一个关键决定\n5. **凡是涉及视觉方案选择，先独立生出候选图，再让用户看图选择**\n6. 每次生图都单独执行，不和解释文字混在同一回复里\n7. 用户选图后更新状态，再进入下一轮细化\n8. 最终询问用户是否还要补充图注、caption、图中文字说明\n\n## 这版 skill 特别强调的协议\n\n### 1. 文字步骤和生图步骤严格分开\n\n正确节奏应该是：\n\n- 第一步：文字回复，说明当前状态，并询问“是否现在生成下一轮候选图”\n- 第二步：独立执行生图动作（Create image / 对应 API）\n- 第三步：文字回复，对已经生成的候选图做编号，然后请用户**根据图来选**\n\n不要把“解释 + 生图 + 让用户选择”揉成一条混合回复。\n\n### 1.5 生图路径必须走 OpenAI Create image / ChatGPT Images 2.0，而不是 SVG\n\n- 框架图候选图和最终图，**不要走 SVG 合成路线**。\n- 在 **ChatGPT web** 中，应走独立的 **Create image** 步骤。\n- 若宿主支持 **Thinking** 或 **Extended Thinking / images with thinking**，优先用这一路径来生成复杂框架图。\n- 在 **Codex / Trae / API 宿主** 中，也应走原生的 ChatGPT 图片生成能力；不要把 mermaid、graphviz、tikz、纯 SVG 输出当成框架图渲染替代方案。\n\n### 2. 让用户选方案时，应优先让用户从图里选\n\n不是：\n\n- 先写一堆方案文字说明\n- 然后让用户凭想象选 A / B / C\n\n而应该是：\n\n- 先简要说明下一批图会比较什么\n- 询问用户是否现在生成这一批候选图\n- 生成多张图\n- 再让用户从图里选择 A / B / C / D\n\n\n### 4. 每次文字回复后，都要告诉用户下面 1–2 步做什么\n\n不要只停在当前轮的说明上。每次文字回复结束时，都应该显式告诉用户：\n\n- 下一步是不是要单独生成下一批图\n- 生成完后用户应该从哪些维度来选图\n- 用户选完之后，再下一步会进入哪一轮细化\n\n推荐固定加一个小结尾：\n\n- **下一步**：是否现在生成这一轮候选图\n- **看图后请你重点反馈**：例如风格、结构、密度、机制解释强度、是否太花、是否太满\n- **然后我会**：更新状态并进入下一轮更细的候选图\n\n### 3. 每一轮开始前都可以问一句\n\n推荐问法：\n\n- “要不要我先生成这一轮的风格候选图？”\n- “要不要我先生成这一轮的结构候选图？”\n- “要不要我先生成下一轮细化图？”\n\n## 推荐的多轮顺序\n\n- 第 0 轮：读取论文内容，整理 Figure Brief\n- 第 1 轮：先选大风格家族，但应通过**风格候选图**来选\n- 第 2 轮：再选结构骨架，但应通过**结构候选图**来选\n- 第 3 轮：再选面向哪类审稿人 + 信息密度，但如果视觉差异明显，也应通过**候选图**来选\n- 第 4 轮：再选人物图标、结果示意、公式多少、对比区大小，并通过**内部视觉语言候选图**来选\n- 第 5 轮：首批综合方案多图生成\n- 第 6 轮：用户选图，归纳喜欢和不喜欢的点\n- 第 7 轮：第二批定向细化生成\n- 第 8 轮：最终定稿 + 询问是否补 caption / legend / panel explanation\n\n## 这个 skill 特别强调的点\n\n- 每次生图前都要有人类确认\n- 每次生成要多张图，不要一开始只出 1 张\n- 提示词要具体，不能只写“学术风”“顶刊风”\n- 图像生成动作必须和普通文字回复分离\n- 凡是视觉方向决策，优先基于**已生成图**来做选择\n- 每轮结束后都要更新状态\n- 最终一定要问用户是否还需要图注和正文配套说明\n\n\n### 1.8 宿主环境提醒必须明确说明\n\n- 如果运行在 **OpenClaw / Codex / Trae / 其他 IDE 或 API 宿主**，框架图生图阶段必须走 **OpenAI ChatGPT Images 2.0**（若宿主提供更高版本，则用更高版本）。\n- 如果该宿主没有可用的 **OpenAI API key**，必须先提醒用户提供或配置 key，再进入生图步骤。\n- 如果运行在 **ChatGPT 网页版**，应要求在 **Extended Thinking** 或宿主当前可用的最强 thinking-assisted 路径下进行，并且**不要让用户手动切换到 Create image 工具模式**；而是由助手在独立的生图步骤中触发原生图片生成。\n- 无论在哪个宿主里，**都不能因为缺少图片能力而改用 SVG**。\n\n## 首次进入 skill 时的提醒\n\n第一次使用这个 skill 时，应该先提醒用户：\n\n- **最佳做法**：先用 `paper-deep-reading` skill 对论文初稿或相关论文生成一份**精读报告**，并保存为 **Markdown**\n- 推荐链接：`https://clawhub.ai/c-narcissus/paper-deep-reading`\n- 但这**不是必须的**\n- 如果用户还没有准备好初稿，也可以先输入：模型模块描述、算法思路、设计思想、算法流程、训练机制、系统结构草图等\n\n然后要继续确认两件事：\n\n1. 用户当前输入是否已经足够开始框架图设计\n2. 用户是否准备好**现在开始画图**\n\n一旦读完精读报告或方法描述，第一次正式设计轮应该明确告诉用户：\n\n- 下一步通常是先生成**一批不同风格的候选图**供选择\n- 这个选择应该基于图，而不是只基于文字描述\n\n\n## 状态记录与后续续接\n\n每次**文字回复**后，除了给用户当前结论，还应更新一份当前状态，至少记录：\n\n- 当前 Figure Brief\n- 已经生成了哪些候选图 / 交付物\n- 用户已经选中了什么、排除了什么\n- 当前正在等待哪一个决定\n- 下一步若用户同意，会生成哪一批图\n\n同时还要提醒用户：为了确保后续在 OpenClaw、Codex、Trae 或其他宿主里能稳定续接，后续每次提问时，最好显式写上类似：\n\n- `请使用 paper-framework-figure-studio-pro 根据当前状态继续执行，并处理下面的新要求：...`\n- `Use paper-framework-figure-studio-pro to continue from the current saved state and apply the following change: ...`\n\n这样能减少宿主环境丢失上下文或没有正确续接状态的风险。\n\n\n## 关键硬约束：文字回复和生图不能同轮出现\n\n- 只要这一轮回复里包含解释、总结、追问、确认、下一步导航，就必须是**纯文字回复**。\n- 如果下一步要生图，这一轮只能先说明将要生成什么，并询问用户是否现在开始生成。\n- 用户确认后，**下一轮 / 下一独立动作**才允许真正调用 OpenAI ChatGPT Images 2.0 / Create image。\n- 绝不能在“说明 + 生图”同一次回答里同时发生。\n\n- 每次文字回复结尾都要提醒用户：如果生成图后不知道下一步该怎么提问，可以直接输入 **`接下来做什么`**，skill 会给出下一步建议和推荐提问方式。\n\n\nRendering rule reminder used in every text-only reply:\n- ChatGPT web: use native image generation under the strongest available thinking-assisted path; prefer Extended Thinking when available; do not ask the user to manually switch to Create image.\n- OpenClaw / Codex / Trae / API hosts: use OpenAI ChatGPT Images 2.0 or newer.\n- SVG, mermaid, tikz, graphviz, and other vector-code fallbacks are forbidden.\n\nFile v1.0.2:references/reviewer_style_taxonomy.md\n\n# Reviewer Style Taxonomy\n\nUse this to tailor the figure direction to the likely reviewer.\n\n## Technical-formal reviewer\nPrefers:\n- cleaner architecture diagrams\n- fewer decorative icons\n- stronger notation discipline\n- moderate or high information density\n\nGood families:\n- Academic Conservative\n- Modern Modular Tiles (formal version)\n- Mechanism + Result Snapshots (formal version)\n\n## Cross-domain reviewer\nPrefers:\n- clearer storytelling\n- stronger intuition cues\n- fewer symbols\n- more obvious problem-to-method narrative\n\nGood families:\n- Modern Modular Tiles\n- Editorial Flat Illustration\n- Mechanism + Result Snapshots\n\n## Design-sensitive reviewer\nPrefers:\n- polished hierarchy\n- modern composition\n- memorable visual system\n- elegance without loss of rigor\n\nGood families:\n- Modern Modular Tiles\n- Editorial Flat Illustration\n- Premium Scientific Illustration\n\nFile v1.0.2:references/visual_communication_principles.md\n\n#\n\nArchive v1.0.1: 21 files, 36729 bytes\n\nFiles: assets/conversation_state.template.json (3988b), assets/figure_brief_template.md (1527b), assets/next_step_navigation.md (4739b), assets/prompt_library.md (11863b), assets/refinement_controls.md (1517b), CHANGELOG.md (2197b), examples/example_opening_turns.md (2407b), publish/cover_and_icon_prompts.md (817b), publish/listing_long.md (1294b), publish/listing_short.md (317b), publish/release_checklist.md (555b), publish/starter_messages.md (1171b), README.md (4825b), references/clawhub_packaging_notes.md (1413b), references/README_CN.md (7495b), references/reviewer_style_taxonomy.md (871b), references/visual_communication_principles.md (1765b), references/workflow_examples.md (8092b), SKILL.md (26003b), templates/user_input_bundle.md (863b), _meta.json (152b)\n\nArchive v1.0.0: 8 files, 7127 bytes\n\nFiles: CHANGELOG.md (316b), docs/CLAWHUB_PUBLISH_COPY.md (2751b), docs/RELEASE_NOTES.md (463b), examples/example-input.md (874b), metadata.json (601b), README.md (1590b), SKILL.md (6118b), _meta.json (152b)","readmeExcerpt":"Skill: Paper Framework Figure Studio Pro Owner: c-narcissus Summary: Use when the user wants to design, prompt, generate, critique, or integrate publication-ready research-paper framework figures: method overview diagrams, arc... Tags: latest:1.2.0 Version history: v1.2.0 | 2026-05-05T12:58:04.848Z | user **Version 1.2.0 – Major workflow and structure upgrade** - Overhauled workflow for strict step-by-step, multi-tur","codeSnippets":[],"executableExamples":[{"language":"text","snippet":"请严格按照 paper-framework-figure-studio-pro-v1.2.0-skill.zip 里 skill 的步骤，对 semiDFL.pdf 绘制 diagram。不要参考 semiDFL.pdf 里面已有的 diagram。"},{"language":"text","snippet":"Please strictly follow the workflow in paper-framework-figure-studio-pro-v1.2.0-skill.zip to draw a diagram for semiDFL.pdf. Do not refer to any existing diagram inside semiDFL.pdf."},{"language":"text","snippet":"Board purpose:\nGenerate candidate images or schematic candidates for choosing a framework-figure direction, not a final figure.\n\nCandidate count:\n6 by default, allowed 4-6.\n\nHold fixed:\n<paper thesis, target slot, required modules, exact labels, color semantics, sample-image transfer rules>\n\nVary only:\n<subtype / scheme / layout / style / metaphor / density / prompt framing>\n\nCompare:\n<what the user should decide by looking at the images>\n\nRendering route:\nChatGPT web: Create image through ChatGPT Images 2.0.\nCodex: $imagegen first; if unavailable, ChatGPT Images 2.0 API or another approved image-generation API."},{"language":"text","snippet":"Create a publication-ready research-paper framework diagram as a raster image.\n\nGoal:\n<figure thesis>\n\nPaper slot and audience:\n<slot and audience>\n\nDiagram subtype and layout:\n<subtype, canvas, panel count, reading order>\n\nRequired content:\n<modules/entities/stages/evidence>\n\nLabels:\nUse only these exact labels: <labels>.\n\nSample-image transfer:\n<per-image transfer rules or \"none\">\n\nStyle and color semantics:\n<style, palette, what colors mean>\n\nCandidate variation:\nGenerate <4/5/6> candidates, usually 6. Vary only <axis>.\n\nAvoid:\nlong paragraphs, microscopic labels, fake metrics, fake UI, logos, watermarks, decorative clutter, SVG/Mermaid/TikZ/Graphviz/code-rendered diagram instructions."}],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"---\nname: paper-framework-figure-studio-pro\nlicense: MIT-0\ndescription: \"Use when the user wants to design, prompt, generate, critique, or integrate publication-ready research-paper framework figures: method overview diagrams, architecture diagrams, pipeline/process diagrams, agent workflows, system/data-flow figures, mechanism-intuition figures, case walkthrough panels, and reviewer-facing schematic figures. Generated from research-paper-figure-skill-factory v1.0.1 with full-feasible local PDF evidence, startup-plan-only first replies, strict text/image separation, mandatory text-candidate to visual-candidate setup to image-only candidate board to candidate-review selection workflow, optional sample images, ChatGPT web Create image / ChatGPT Images 2.0, Codex $imagegen first, all-step/current-position state footers, and next-question help in every text reply.\"\nmetadata:\n  display_name: Paper Framework Figure Studio Pro\n  version: \"1.2.0\"\n  author: OpenAI\n  tags:\n    - research-figure\n    - paper-framework\n    - method-framework\n    - architecture-diagram\n    - pipeline-diagram\n    - agent-workflow\n    - candidate-image-bridge\n    - imagegen\n    - chatgpt-images-2\n    - clawhub\n    - openclaw\n  compatibility: Codex, ChatGPT web, OpenClaw, ClawHub marketplace. Requires image-generation capability for rendering.\n  openclaw:\n    skillKey: paper-framework-figure-studio-pro\n---\n\n# Paper Framework Figure Studio Pro\n\nThis skill designs publication-ready raster framework figures for computer-science research papers. Use it for method overviews, architecture diagrams, pipelines, agent workflows, system/data-flow figures, mechanism-intuition figures, case walkthroughs, and reviewer-facing schematic figures.\n\nIt was regenerated with `research-paper-figure-skill-factory` v1.0.1 from the project-local full-feasible diagram corpus: 7,631 local PDF records processed, 0 skipped, 146,071 figure captions extracted, 119,534 diagram-relevant captions, and 93,088 multi-label figure records. Framework-relevant evidence includes method-framework, architecture, pipeline/process, agent-workflow, mechanism, and case-walkthrough patterns. Representative rendered pages are audit aids only, not the corpus size.\n\n## Non-Negotiable Contract\n\n### First Trigger\n\nOn the first reply in a new project, output only a startup plan. Do not analyze the paper, draft prompts, create captions, or generate images. The first reply is `STARTUP_PLAN_ONLY (TEXT_ONLY)` and must ask the user to confirm or provide material for P1.\n\nIf the first user message asks to \"直接出图\", \"生成 6 张图\", \"出候选图\", \"generate images\", or otherwise asks for image generation, record the request as pending only. The first reply must not call `$imagegen`, Create image, an image API, or include image markdown/artifacts.\n\n### Mandatory Candidate-Image Bridge\n\nAfter any multi-option text decision, do not move directly to final prompt, final image generation, caption, or text-only locking. Use this mandatory bridge:\n\n1. `TEXT_ONL"},{"path":"README.md","content":"# Paper Framework Figure Studio Pro\n\n## 中文\n\n`paper-framework-figure-studio-pro` 用于帮助研究者为论文生成 framework diagram、method diagram、pipeline diagram、architecture diagram 和 agent workflow 等框架图。它适合把论文 PDF、摘要、方法说明或草稿转化为可比较的制图方案、候选图、修改建议、caption 和图注说明。\n\n### 推荐使用方式\n\n优先在 ChatGPT 网页版中使用，并选择 **Extended thinking**。网页版更适合完成完整的论文理解、候选图生成和多轮修图流程。\n\n如果接下来的步骤是生成图片，最好在 ChatGPT 网页版中手动选择 **Create image** 模式，再让它继续生成候选图或最终图。\n\n在 Codex 里也可以尝试使用，但可能会遇到图像生成或上下文处理问题，而且会比较费 token。除非你明确需要在本地工程目录中整理文件、改 skill 或生成配套文档，否则不建议把主要制图流程放在 Codex 里完成。\n\n### ChatGPT 网页版使用步骤\n\n1. 把 `paper-framework-figure-studio-pro-v1.2.0-skill.zip` 放进 ChatGPT 的 Sources。\n2. 把论文 PDF 也放进 Sources，例如 `semiDFL.pdf`。\n3. 选择 Extended thinking。\n4. 输入类似下面的 prompt：\n\n```text\n请严格按照 paper-framework-figure-studio-pro-v1.2.0-skill.zip 里 skill 的步骤，对 semiDFL.pdf 绘制 diagram。不要参考 semiDFL.pdf 里面已有的 diagram。\n```\n\n如果你的论文文件名不是 `semiDFL.pdf`，请把 prompt 里的文件名替换为实际上传到 Sources 的文件名。\n\n当 skill 已经完成文字方案比较，并提示下一步要生成候选图或最终图时，建议手动切换到 **Create image** 模式后再继续。\n\n### 制图流程\n\n1. 提供论文 PDF、摘要、方法说明、目标章节或已有草稿。\n2. 说明是否要避开论文中已有 diagram，以及是否提供参考图。\n3. Skill 先判断这张图更适合 framework、architecture、pipeline、workflow 还是 mechanism diagram。\n4. 先生成 4-6 个文字方案，通常是 6 个。\n5. 选择或确认候选图方向；如果有参考图，可以说明每张图只参考布局、风格、信息密度、标签或配色中的哪些属性。\n6. 生成多张候选图或示意图供比较。\n7. 从候选图中选择最接近的一张，或指出需要修改的地方。\n8. 根据选择继续生成正式版本或修订版本。\n9. 最后整理 caption、legend 和正文中的图说明文字。\n\n## English\n\n`paper-framework-figure-studio-pro` helps researchers create framework diagrams, method diagrams, pipeline diagrams, architecture diagrams, and agent workflows for research papers. It turns a paper PDF, abstract, method description, or draft notes into comparable diagram directions, candidate figures, revision guidance, captions, and figure descriptions.\n\n### Recommended Use\n\nPrefer using this skill in the ChatGPT web app with **Extended thinking** enabled. The web app is better suited for the full workflow: paper understanding, candidate figure generation, and iterative figure revision.\n\nIf the next step is image generation, it is best to manually select **Create image** mode in the ChatGPT web app before asking it to generate candidate figures or the final figure.\n\nYou can also try it in Codex, but image generation and context handling may be less reliable, and it can consume many tokens. Unless you specifically need local file organization, skill editing, or repository documentation, the main figure-making workflow is better done in ChatGPT web.\n\n### ChatGPT Web Usage\n\n1. Add `paper-framework-figure-studio-pro-v1.2.0-skill.zip` to ChatGPT Sources.\n2. Add the paper PDF to Sources as well, for example `semiDFL.pdf`.\n3. Select Extended thinking.\n4. Type a prompt like this:\n\n```text\nPlease strictly follow the workflow in paper-framework-figure-studio-pro-v1.2.0-skill.zip to draw a diagram for semiDFL.pdf. Do not refer to any existing diagram inside semiDFL.pdf.\n```\n\nIf your paper file is not named `semiDFL.pdf`, replace the file name in the prompt with the exact file name uploaded to Sources.\n\nWhen the skill has finished comparing text di"},{"path":"_meta.json","content":"{\n  \"ownerId\": \"kn7fxns1xpr6z67w885my7d7k98506vv\",\n  \"slug\": \"paper-framework-figure-studio-pro\",\n  \"version\": \"1.2.0\",\n  \"publishedAt\": 1777985884848\n}"},{"path":"references/builder-time-acquisition-report.md","content":"# Builder-Time Acquisition Report\n\nGenerated by `research-paper-figure-skill-factory` v1.0.1 for `paper-framework-figure-studio-pro` v1.2.0.\n\n## Source\n\nThe skill uses the previously extracted project-local full-feasible diagram corpus, rooted at `research-paper-diagram-generation-corpus/`, as builder-time evidence. No new web download was needed for this generation because the local PDF index and extraction artifacts already existed.\n\n## Processing Scope\n\n- Scope: `all_accessible_relevant_pdfs`.\n- Candidate PDFs: 7,631.\n- Accessible PDFs: 7,631.\n- Processed PDFs: 7,631.\n- Skipped PDFs: 0.\n- Skipped reasons: none recorded.\n- Representative rendered pages: 96 audit aids, not the corpus size.\n\n## Framework Evidence Focus\n\nThe generated skill narrows the broader diagram taxonomy to framework-related figure production: method frameworks, architecture diagrams, pipeline/process diagrams, agent workflows, system/data-flow diagrams, mechanism-intuition figures, and case walkthroughs.\n\nProduction-grade lock is supported because the skill includes source-corpus notes, an evidence-map index, an evidence-lineage summary, and framework-specific taxonomy/pattern references."},{"path":"references/evidence-lineage-summary.md","content":"# Evidence Lineage Summary\n\nThe full local evidence map supports these skill claims:\n\n1. Framework-diagram routing must distinguish method framework, architecture, pipeline/process, agent workflow, system/data flow, graph/network, mechanism, walkthrough, evidence-board, taxonomy, data/protocol, failure, and theory/proof-intuition layouts.\n2. Reader questions differ across diagrams: system identity, process sequence, entity relation, mechanism, case behavior, contribution boundary, and claim support.\n3. Density must be selected before rendering because framework figures range from clean overview panels to dense evidence-linked boards.\n4. Panel choreography is a first-class design decision for multi-panel research figures.\n5. Evidence diagrams must not invent results; they require user-provided metrics, comparisons, or qualitative examples.\n6. Routing must be multi-label: one PDF or one diagram can support multiple labels before a primary production subtype is selected.\n7. When multiple schemes are plausible, the skill should move toward generated candidate images or schematic candidates, usually 6, instead of asking the user to compare only text.\n\nThese claims are backed by `research-paper-diagram-generation-corpus/extracted/evidence_map.json`.\n\n## Full-Corpus Counts\n\n- Processed local PDF records: 7,631.\n- Verified official oral PDFs: 3,356.\n- Supplemental local PDFs: 4,275.\n- Extracted figure captions: 146,071.\n- Diagram-relevant captions: 119,534.\n- Multi-label diagram records: 93,088.\n- Representative rendered pages: 96.\n\n## Framework-Relevant Label Coverage\n\nThe label counts are multi-label and can sum above the number of diagram-relevant captions.\n\n| Label | Caption count | Paper count | Representative pages |\n|---|---:|---:|---:|\n| method_framework | 16,565 | 5,267 | 8 |\n| architecture | 14,651 | 4,256 | 8 |\n| pipeline_process | 28,981 | 5,878 | 8 |\n| agent_workflow | 16,271 | 2,228 | 8 |\n| graph_network | 26,059 | 4,335 | 8 |\n| mechanism_intuition | 42,578 | 5,947 | 8 |\n| case_walkthrough | 35,822 | 6,465 | 8 |\n| evidence_board | 63,404 | 6,947 | 8 |\n| data_benchmark_protocol | 25,583 | 4,881 | 8 |\n| failure_limitation | 20,719 | 4,570 | 8 |\n| taxonomy_design_space | 5,993 | 2,141 | 8 |\n| theory_proof_intuition | 9,985 | 3,286 | 8 |\n| general_diagram_or_figure | 17,072 | 3,160 | 0 |"}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":null,"editorialQuality":{"score":100,"threshold":65,"status":"thin","wordCount":2078,"uniquenessScore":41,"reasons":["uniqueness-below-45"]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-10T15:27:10.104Z","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":"2026-10-10T15:27:10.104Z","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":"2026-10-10T17:35:53.061Z","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"}]}}}