{"id":"55bd9a3c-a1f2-40d4-897f-5c3123659947","entityType":"agent","slug":"clawhub-tinadu-ai-deck-pipeline","name":"Deck Pipeline","canonicalUrl":"https://www.xpersona.co/agent/clawhub-tinadu-ai-deck-pipeline","canonicalPath":"/agent/clawhub-tinadu-ai-deck-pipeline","generatedAt":"2026-10-11T14:15:22.201Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-11T10:26:17.577Z","emptyReason":null},"description":"Production-grade Claude Code system that takes presentation decks from raw Chinese draft to McKinsey-polished English — with full audit trail, layout integri... Skill: Deck Pipeline Owner: tinadu-ai Summary: Production-grade Claude Code system that takes presentation decks from raw Chinese draft to McKinsey-polished English — with full audit trail, layout integri... Tags: bilingual:3.0.0, claude-code:3.0.0, latest:3.0.0, layout-audit:3.0.0, mckinsey:3.0.0, pptx:3.0.0, translation:3.0.0 Version history: v3.0.0 | 2026-05-12T16:33:27.011Z | user Major rewrite from DeckGlobalize","descriptionLabel":"Technical summary","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 1.1K downloads reported by the source. Last updated 10/11/2026.","installCommand":"clawhub skill install s178xc11mj68f9c018344m5xw183y7z8:deck-pipeline","sourceUrl":"https://clawhub.ai/tinadu-ai/deck-pipeline","homepage":"https://clawhub.ai/tinadu-ai/skills/deck-pipeline","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/tinadu-ai/deck-pipeline","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/tinadu-ai/skills/deck-pipeline","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":61,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Production-grade Claude Code system that takes presentation decks from raw Chinese draft to McKinsey-polished English — with full audit trail, layout integri..."},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-11T10:26:17.577Z","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":"2026-10-11T10:26:17.577Z","emptyReason":null},"stars":null,"forks":null,"downloads":1088,"packageName":null,"latestVersion":"3.0.0","tractionLabel":"1.1K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-11T10:26:17.510Z","emptyReason":null},"lastUpdatedAt":"2026-10-11T10:26:17.577Z","lastCrawledAt":"2026-10-11T10:26:17.510Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-12T10:26:17.510Z","lastVerifiedAt":null,"highlights":[{"version":"3.0.0","createdAt":"2026-05-12T16:33:27.011Z","changelog":"Major rewrite from DeckGlobalizer v2.1.1: - Now a full 4-stage pipeline (Sense Pass → McKinsey Translation → Layout Audit → Handoff) - Swappable PROFILE block (palette, fonts, glossary, slide-offset) - Page-by-page execution with per-slide checkpoint - McKinsey style baseline + extensible reference samples - Font-name suffix audit, structural-anchor detection - Discrete -0.1pt font compression, line-spacing fallback - Bidirectional PPT ↔ Excel sync - Companion-file lock detection - 10 standalone helper scripts - Polish-only mode for single-language decks Renamed from 'deckglobalizer' to 'deck-pipeline'; old slug redirects.","fileCount":17,"zipByteSize":43050},{"version":"2.1.1","createdAt":"2026-04-27T08:01:49.827Z","changelog":"Add requires field declaring python-pptx and lxml dependencies. Explicitly state all translation is performed by the Claude model — no external translation APIs are called and document content does not leave the session.","fileCount":2,"zipByteSize":4703},{"version":"2.1.0","createdAt":"2026-04-27T07:53:20.798Z","changelog":"Clarify layout compensator: refining translation must not drop factual content, numbers, or named concepts. Remove cross-file slide copying section.","fileCount":2,"zipByteSize":4546},{"version":"1.0.0","createdAt":"2026-04-22T05:02:32.467Z","changelog":"- Initial release of DeckGlobalizer skill for high-fidelity PowerPoint translation. - Translates .pptx files into another language while preserving original layout, fonts, and design. - Multi-phase workflow: Style audit with manifest, tiered glossary with user sign-off, page-by-page translation with layout compensator rules. - Strict text fitting and sibling-consistency logic to maintain professional appearance. - Outputs translated deck and intermediate documentation files in the source file directory.","fileCount":2,"zipByteSize":2858}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s178xc11mj68f9c018344m5xw183y7z8:deck-pipeline","setupComplexity":"low","setupSteps":["Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.","Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data."],"contract":{"contractStatus":"missing","authModes":[],"requires":[],"forbidden":[],"supportsMcp":false,"supportsA2a":false,"supportsStreaming":false,"inputSchemaRef":null,"outputSchemaRef":null,"dataRegion":null,"contractUpdatedAt":null,"sourceUpdatedAt":null,"freshnessSeconds":null},"invocationGuide":{"preferredApi":{"snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-11T14:15:22.200Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tinadu-ai-deck-pipeline/trust"}},"reliability":{"evidence":{"source":"runtime-metrics","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No trust, reliability, or runtime telemetry is available."},"trust":{"status":"unavailable","handshakeStatus":"UNKNOWN","verificationFreshnessHours":null,"reputationScore":null,"p95LatencyMs":null,"successRate30d":null,"fallbackRate":null,"attempts30d":null,"trustUpdatedAt":null,"trustConfidence":"unknown","sourceUpdatedAt":null,"freshnessSeconds":null},"decisionGuardrails":{"doNotUseIf":["Contract metadata is missing or unavailable for deterministic execution."],"safeUseWhen":[],"riskFlags":["missing_or_unavailable_contract","trust_data_unavailable","schema_references_missing"],"operationalConfidence":"low"},"executionMetrics":{"observedLatencyMsP50":null,"observedLatencyMsP95":null,"estimatedCostUsd":null,"uptime30d":null,"rateLimitRpm":null,"rateLimitBurst":null,"lastVerifiedAt":null,"verificationSource":null},"runtimeMetrics":{"successRate":null,"avgLatencyMs":null,"avgCostUsd":null,"hallucinationRate":null,"retryRate":null,"disputeRate":null,"p50Latency":null,"p95Latency":null,"lastUpdated":null}},"benchmarks":{"evidence":{"source":"no-benchmark-data","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No benchmark suites or observed failure patterns are available."},"suites":[],"failurePatterns":[]},"artifacts":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"high","updatedAt":"2026-10-11T10:26:17.577Z","emptyReason":null},"readme":"Skill: Deck Pipeline\n\nOwner: tinadu-ai\n\nSummary: Production-grade Claude Code system that takes presentation decks from raw Chinese draft to McKinsey-polished English — with full audit trail, layout integri...\n\nTags: bilingual:3.0.0, claude-code:3.0.0, latest:3.0.0, layout-audit:3.0.0, mckinsey:3.0.0, pptx:3.0.0, translation:3.0.0\n\nVersion history:\n\nv3.0.0 | 2026-05-12T16:33:27.011Z | user\n\nMajor rewrite from DeckGlobalizer v2.1.1:\n\n- Now a full 4-stage pipeline (Sense Pass → McKinsey Translation → Layout Audit → Handoff)\n- Swappable PROFILE block (palette, fonts, glossary, slide-offset)\n- Page-by-page execution with per-slide checkpoint\n- McKinsey style baseline + extensible reference samples\n- Font-name suffix audit, structural-anchor detection\n- Discrete -0.1pt font compression, line-spacing fallback\n- Bidirectional PPT ↔ Excel sync\n- Companion-file lock detection\n- 10 standalone helper scripts\n- Polish-only mode for single-language decks\n\nRenamed from 'deckglobalizer' to 'deck-pipeline'; old slug redirects.\n\nv2.1.1 | 2026-04-27T08:01:49.827Z | user\n\nAdd requires field declaring python-pptx and lxml dependencies. Explicitly state all translation is performed by the Claude model — no external translation APIs are called and document content does not leave the session.\n\nv2.1.0 | 2026-04-27T07:53:20.798Z | user\n\nClarify layout compensator: refining translation must not drop factual content, numbers, or named concepts. Remove cross-file slide copying section.\n\nv1.0.0 | 2026-04-22T05:02:32.467Z | auto\n\n- Initial release of DeckGlobalizer skill for high-fidelity PowerPoint translation.\n- Translates .pptx files into another language while preserving original layout, fonts, and design.\n- Multi-phase workflow: Style audit with manifest, tiered glossary with user sign-off, page-by-page translation with layout compensator rules.\n- Strict text fitting and sibling-consistency logic to maintain professional appearance.\n- Outputs translated deck and intermediate documentation files in the source file directory.\n\nArchive index:\n\nArchive v3.0.0: 17 files, 43050 bytes\n\nFiles: CHANGELOG.md (2306b), CONTRIBUTING.md (2162b), LICENSE (1083b), README.md (16402b), scripts/anchor_detect.py (3582b), scripts/apply.py (5166b), scripts/excel_sync.py (10181b), scripts/extract.py (2096b), scripts/glossary_audit.py (4945b), scripts/handoff.py (2260b), scripts/layout_audit.py (7996b), scripts/overflow_recheck.py (6081b), scripts/sense_pass.py (6505b), scripts/style_distill.py (6272b), skill-card.md (2294b), SKILL.md (17289b), _meta.json (132b)\n\nFile v3.0.0:SKILL.md\n\n---\nname: deck-pipeline\ndescription: |\n Production-grade Claude Code system that takes presentation decks from raw\n Chinese draft to McKinsey-polished English — with full audit trail, layout\n integrity checks, and a swappable PROFILE block for project-specific\n defaults. Built on a 4-stage pipeline (Sense Pass → McKinsey Translation →\n Layout Audit → Handoff). Also runs polish-only on any single-language deck.\n\n TRIGGER when the user:\n • hands over a .pptx containing Chinese and asks for English / translation\n • asks for \"deck pipeline\", \"deck polish\", \"deck globalizer\"\n • asks for layout polish, font cleanup, overflow fixing on any deck\n • asks to update / reverse-sync a bilingual comparison Excel against a deck\n\n SUPPRESS with \"Ignore deck-pipeline\".\nversion: 0.1.0\nlicense: MIT\n---\n\n# Deck Pipeline\n\n> **A 4-stage, audit-trailed Chinese→English deck globalization system with a swappable project profile.**\n\nThis skill bundles the generic deck-globalization engine (originally upstream\nDeckGlobalizer v2.1.1) and an editable PROFILE block (palette, fonts,\nglossary, style preferences). The two are **separated** by section so the\nprofile can be swapped per project / brand without touching the engine.\n\nFor a marketing-style overview, see `README.md` in this directory.\nFor implementation, see `scripts/` and the per-stage runbooks below.\n\n---\n\n## 0. Modes & activation\n\n| Mode | Trigger | Stages |\n|---|---|---|\n| **Full pipeline** | CN deck (± EN draft) + user wants English output | 1 → 2 → 3 → 4 |\n| **Polish-only** | Single-language deck + \"layout / format only / skip translation\" | 1 → 3 → 4 |\n| **Reverse-sync only** | User hand-edited a PPT after a comparison Excel was generated | 3.5 (sync sub-routine) |\n\nDetect the mode in the first turn. If ambiguous, ask one yes/no question\n(\"This deck is already in EN — should I just polish layout, or also rewrite\nMcKinsey-style?\"). Do not guess silently.\n\n---\n\n## 1. PROFILE block — defaults (swappable)\n\nEdit this block to retarget the skill for your project / brand. Everything\nbelow this block is **profile-agnostic**.\n\n```yaml\nPROFILE:\n # ---- L1 Tokens ----\n palette:\n # Replace with your brand colors.\n ink: \"#1A1A1A\"\n primary: \"#000000\" # accent / brand primary\n soft: \"#FFFFFF\" # soft fill behind banners\n page_bg: \"#FFFFFF\"\n fonts:\n # Choose a serif title face + a sans-serif body face for best contrast.\n title: \"Georgia\"\n body: \"Verdana\"\n title_bold: true\n unit_table:\n # Chinese number magnitudes → English. 亿 is 100M, NOT \"billion\".\n \"百万\": \"M\"\n \"千万\": \"10M\"\n \"亿\": \"100M\"\n \"十亿\": \"1B\"\n \"百亿\": \"10B\"\n \"千亿\": \"100B\"\n \"万亿\": \"1T\"\n # Currency suffix is left to the user — append \"$\" / \"RMB\" / \"€\" as appropriate.\n\n # ---- L2 Constants ----\n size_ladder: [22, 14, 10, 8, 6, 4] # H1, H2, body, caption, footnote, source\n floors:\n body: 7\n caption: 6\n source: 4\n compression_step: 0.1 # discrete -0.1pt iterations only\n line_height_default: 1.25\n line_height_fallback: 1.15 # used before sub-floor compression\n quote_style: \"single\" # 'McKinsey' single quotes\n footer_format: \"Confidential · For Intended Recipients Only · {month} {year}\"\n separator_in_footer: \"·\" # middle dot, NOT em-dash\n\n # ---- L1 Glossary (extensible) ----\n # Replace the example entries below with your project's locked terms.\n # Categories are illustrative; you can rename / add / remove.\n glossary:\n locked:\n people_orgs:\n # \"\": \"\"\n # e.g. \"John Smith\": \"John Smith\"\n # e.g. \"Acme Capital\": \"Acme Capital\"\n {}\n business_terms:\n # Common Chinese business-deck idioms with industry-standard\n # English mappings. Edit / extend as needed.\n \"流水\": \"gross revenue\"\n \"私域\": \"owned audience\"\n \"出海\": \"global expansion\"\n domain_specific:\n # Project / industry / domain terms.\n # \"\": \"\"\n {}\n rejected_rewrites:\n # Entries the user vetoed during prior sessions.\n # Format: { source: \"...\", proposed: \"...\", reason: \"...\" }\n []\n pending: []\n session_added: []\n\n # ---- Style rules ----\n # McKinsey is the default baseline. Additional style references can be\n # uploaded and distilled via scripts/style_distill.py; their rules layer\n # ON TOP of the McKinsey base.\n style_baseline: \"mckinsey\"\n mckinsey:\n title_is_takeaway: true # title = the so-what, not the topic\n lead_with_so_what: true\n parallel_structure: true # bullets share tense, opening part-of-speech\n strong_action_verbs: true # cut \"is/has\", prefer concrete verb\n cut_filler:\n - \"in order to → to\"\n - \"a number of → many\"\n - \"due to the fact that → because\"\n - \"at this point in time → now\"\n case: \"sentence\" # lowercase unless proper noun or locked term\n em_dash_policy: \"use em-dash for parentheticals; use · (middle dot) in lists/footers\"\n style_references:\n # Each entry is a PDF / .pptx reference. style_distill.py reads it and\n # emits rules (cadence, signature phrases, paragraph length, tone) that\n # layer on top of the McKinsey base. Conflicts: more recent entry wins;\n # user is asked at first conflict.\n # Example:\n # - path: \"/path/to/sample.pdf\"\n # weight: 0.7\n []\n\n # ---- Structural anchor heuristics ----\n anchor_detection:\n min_pages: 3 # appears on ≥3 slides\n match_on: # signature components\n - position_xy\n - fill_color\n - font_size_class\n auto_protect: true\n\n # ---- Overflow estimator ----\n overflow:\n severity:\n high: 1.5\n med: 1.15\n low: 1.0\n surface_only: \"high\" # surface MED/LOW only when explicitly asked\n defer_to_user_threshold: 10 # if HIGH > 10 → ask user to render externally\n\n # ---- CN ↔ EN slide alignment ----\n # Default is 1:1 (EN slide N maps to CN slide N).\n # Set overrides only when the two decks have been restructured.\n # Pass this config to excel_sync.py via `--cn-offset `.\n cn_en_slide_offset:\n default: 0 # offset added to EN slide number (0 = 1:1)\n overrides: {} # e.g. {\"9-26\": -1, \"20\": null}\n # int = relative offset; null = no CN counterpart\n```\n\n> **Profile-agnostic note:** all sections below treat `PROFILE` as an\n> opaque dict. Do not hardcode project-specific values anywhere outside the\n> PROFILE block.\n\n---\n\n## 2. Pipeline stages\n\nEach stage has: **inputs · what it does · outputs · stop-and-ask conditions.**\n\n### Stage 1 — Sense Pass\n\n**Inputs:** one or two `.pptx` paths (CN, optional EN draft)\n**What it does:**\n1. Run `scripts/sense_pass.py` to extract:\n - palette (top fill colors)\n - font usage histogram\n - size distribution\n - title-zone shapes (top ≤ 600K EMU)\n - layout heuristics\n2. Cross-check sensed values against `PROFILE.palette` / `PROFILE.fonts`.\n If a sensed font is NOT in the whitelist AND NOT in `SKIP_POLLUTION`,\n record it as **font pollution**.\n3. Surface **candidate glossary entries**: any CN noun phrase that appears\n ≥2 times and isn't already in `glossary.locked`.\n\n**Outputs:**\n- `Style_Manifest.md` (in-memory; not written to disk unless requested)\n- `pollution_report` (slide → font → count)\n- `candidate_glossary` (term → count → sample context)\n\n**Stop-and-ask:**\n- Candidate glossary surfaces a term Claude can't confidently translate →\n ask user, write answer to `glossary.session_added`.\n- Sensed primary palette color differs from `PROFILE.palette.primary` →\n ask whether to update profile or keep existing.\n\n---\n\n### Stage 2 — McKinsey Translation (skipped in polish-only mode)\n\n**Inputs:** Stage 1 outputs + the CN deck (and optional EN draft for diff context) + any uploaded `style_references`.\n\n**Style layering**: McKinsey base rules (`PROFILE.mckinsey`) apply first. If\n`PROFILE.style_references` is non-empty, run `scripts/style_distill.py` on\neach reference before translation begins; the distilled rules (cadence,\nsignature phrases, paragraph length, tone) layer on top. More recent entry\nwins on conflict; ask user at the first conflict.\n\n**Page-by-page execution (hard requirement):**\n1. **Overall confirmation first** — after Stage 1, show the user the planned\n per-page edit count + sample of style rules in effect; wait for \"go\".\n2. **Then loop slides 1 → N**, one at a time:\n - Collect paragraph-level CN text on this slide via `scripts/extract.py`.\n - For each paragraph, produce EN per the layered style rules:\n - lowercase by default; title = so-what; parallel bullets; strong verbs;\n filler-word table applied; glossary `locked` inline\n - any unknown term → STOP, ask user, write to `session_added`\n - Build the slide's edit batch as a JSON object.\n - Run `scripts/apply.py` with the slide's batch → writes that slide's\n changes into `-en-polished-.pptx` AND appends rows to\n `-bilingual-diff-.xlsx` immediately.\n - **Checkpoint**: print \"P{n} done — N changes applied. Continue?\" and\n wait for user OK before moving to P{n+1}.\n - User can interject \"back to P{n-1}\" or \"stop here\" between pages.\n\n**Why per-page (not all-at-once):**\n- The user can review and steer mid-stream.\n- A bad assumption on P3 doesn't propagate to P27 unnoticed.\n- Excel grows incrementally — survives any mid-session interruption.\n- Token-efficient: only one slide's context in active scratchpad.\n\n---\n\n### Stage 3 — Layout Audit\n\n**Inputs:** the post-translation deck (or, in polish-only mode, the raw deck)\n**What it does:**\n\n#### 3a. Font pollution cleanup\nRun `scripts/layout_audit.py --fix`:\n- For every run whose `font.name` is NOT in the title/body whitelist\n OR ends in a style suffix (`Bold` / `Regular` / `Italic` / `Light`):\n - Strip the suffix\n - Set `font.name` to the pure family\n - Set `font.bold` / `font.italic` attributes accordingly\n- Skip any face in the configured `SKIP_POLLUTION` set.\n\n#### 3b. Structural-anchor detection\nRun `scripts/anchor_detect.py`:\n- For each shape, compute a signature: `(rounded_position, fill_color, font_size_class)`.\n- Group across slides. Any signature occurring on ≥ `PROFILE.anchor_detection.min_pages`\n pages becomes an **anchor**.\n- Build `per_page_protect[page] = [anchor_shape_ids...]`.\n- Surface the anchor list to the user. They can add/remove.\n\n#### 3c. Overflow estimation\nRun `scripts/overflow_recheck.py`:\n- Honor `auto_size` (skip if SHAPE_TO_FIT_TEXT or TEXT_TO_FIT_SHAPE).\n- Read actual `margin_*`.\n- Use `PROFILE.line_height_default = 1.25` initially. If a shape is flagged,\n try 1.15 as a what-if before flagging as HIGH.\n- Per-character width by class (narrow `iIl`, wide `MW`, digits, upper, space).\n- Greedy word-wrap simulation.\n- Emit only HIGH (`ratio > PROFILE.overflow.severity.high`) by default.\n\nIf HIGH count > `PROFILE.overflow.defer_to_user_threshold`:\n- **Do not** dump 30+ rendered PNGs into the session.\n- Tell the user: \"Render to PDF/PNG via Keynote or PowerPoint, tell me which\n pages look broken, I'll fix those targeted pages.\"\n\n#### 3d. Compression (when user OKs a fix)\nFor each shape needing fix:\n1. Is it in `per_page_protect[page]`? → SKIP (it's an anchor).\n2. Try widening: increase shape `width` until ratio < 1.0 OR shape collides.\n3. Still > 1.0? Try line-height 1.25 → 1.15.\n4. Still > 1.0? Iterate `font.size -= PROFILE.compression_step` (0.1pt) until\n floor (`PROFILE.floors.`) hit.\n5. Still > 1.0 at floor? **STOP. Escalate to user.** List the shape, its\n current size, the calculated ratio, and ask whether to break the floor.\n\n#### 3e. Late-stage glossary re-scan\nRun `scripts/glossary_audit.py`:\n- For each text run in the deck, check against `glossary.locked`:\n - If a CN-side phrase exists locked but a non-canonical EN translation\n appears → flag.\n - If the same source term is translated two different ways in the deck\n (wavering) → flag.\n- Surface flagged rows. Auto-fix if all flags point to the same canonical\n translation; ask otherwise.\n\n#### 3f. Reverse sync (sub-routine, also Mode 3.5 entry point)\nRun `scripts/excel_sync.py --reverse`:\n- Diff current PPT against the Excel's `en_optimized` column.\n- For each mismatched row:\n - Try ordinal-position match (slide + paragraph-index).\n - If no match, try `difflib.get_close_matches` against same-slide texts.\n - Update Excel cell on success.\n- Report any leftover unmatched rows.\n\n**Outputs:**\n- `-final-.pptx` (full pipeline) or `-final-.pptx` (polish-only)\n- Updated Excel (if applicable)\n\n---\n\n### Stage 4 — Handoff\n\n**Inputs:** all prior-stage outputs\n**What it does:**\n1. Write `HANDOFF.md` to the same directory as the deck — see `scripts/handoff.py`.\n2. Print a one-paragraph deliverables summary to the user.\n\n**Stop-and-ask:** none.\n\n---\n\n## 3. Operational rules (apply across stages)\n\n### 3.1 File-write discipline\n\nBefore writing **any** `.pptx` or `.xlsx`:\n1. Check for `~$` lock file in the same directory.\n2. If present → **STOP.** Tell the user: \"`` is open in\n PowerPoint/Excel. Save and close it, then say 'go' to continue.\"\n3. After writing, immediately readback-verify (next rule).\n\n### 3.2 Excel companion three guard-rails\n\n1. **Pre-write check** — load existing Excel (if any), confirm header row is\n `[page, kind, cn, en_original, en_optimized, notes]`. If columns missing\n → rebuild header before writing data.\n2. **Post-write readback** — immediately reload the saved file and assert\n `max_column ≥ 7` and header is intact.\n3. **Reverse sync** available on demand: see Stage 3f.\n\n### 3.3 Font compression discipline\n\nSee Stage 3d. The single rule: **never** bulk-reduce font sizes.\nAlways discrete `-0.1pt`, always after exhausting widening + line-height\nfallback, always with anchor protection.\n\n### 3.4 Glossary discipline\n\n- Ask once per session per unknown term. Then it's in `session_added` for\n the rest of the session.\n- At handoff, promote `session_added` to a `glossary_proposed_additions.yaml`\n file next to the deck. The user can copy them into PROFILE for the next run.\n- **Never** silently apply a translation Claude is unsure about. Stop and ask.\n\n### 3.5 Magnitude verification\n\nAny number with a CN magnitude word (百万 / 千万 / 亿 / 百亿 / 千亿 / 万亿)\nmust be re-verified against `PROFILE.unit_table` before being written to EN.\nTreat this as a HARD CHECK; do NOT take prior-session translations on faith.\n\n### 3.6 CN-alignment confidence\n\nWhen auto-aligning the Excel's `cn` column by paragraph ordinal:\n- Slides with > 15 changes → auto-tag `notes` column as `needs-review`.\n- Always present this as best-effort, never as ground truth.\n\n---\n\n## 4. Scripts (in `scripts/`)\n\n| Script | Role | Stage |\n|---|---|---|\n| `sense_pass.py` | extract design DNA, font usage, palette | 1 |\n| `extract.py` | paragraph-level text extraction | 1, 2, 3 |\n| `apply.py` | apply EN edits + write Excel with highlight | 2 |\n| `layout_audit.py` | font pollution cleanup, suffix audit | 3a |\n| `anchor_detect.py` | cross-page anchor signature detection | 3b |\n| `overflow_recheck.py` | overflow estimator with severity tiers | 3c |\n| `glossary_audit.py` | late-stage glossary re-scan + wavering | 3e |\n| `excel_sync.py` | bidirectional PPT ↔ Excel sync (configurable slide offset) | 3f |\n| `handoff.py` | write HANDOFF.md | 4 |\n| `style_distill.py` | distill style fingerprint from a reference PDF/.pptx | pre-2 |\n\nEach script is invokable standalone; the skill wires them together.\n\n---\n\n## 5. Deliverables (recap)\n\n**Full pipeline (4 files):**\n- `-en-polished-.pptx`\n- `-final-.pptx`\n- `-bilingual-diff-.xlsx`\n- `HANDOFF.md`\n\n**Polish-only (3 files):**\n- `-final-.pptx`\n- `-layout-changes-.xlsx`\n- `HANDOFF.md`\n\n**Mode 3.5 (reverse-sync only):**\n- updated `-bilingual-diff-.xlsx`\n\n---\n\n## 6. Known limitations\n\n1. Overflow estimator is a **hint, not a verdict** — final visual check\n requires external rendering (Keynote / PowerPoint export to PDF).\n2. `python-pptx` cannot render slides. There is no built-in preview.\n3. CN auto-alignment by paragraph ordinal can drift on heavily-restructured\n pages — configure `PROFILE.cn_en_slide_offset.overrides` for known cases.\n4. The skill assumes the CN source is semantic ground truth — typos in CN\n propagate to EN unless the user catches them.\n5. File-lock collisions silently corrupt output. The pre-write `~$xxx` check\n is the only line of defense.\n\n---\n\n## 7. Changelog\n\nSee `CHANGELOG.md`.\n\n---\n\n## 8. Credits\n\nGeneric deck-globalization engine derived from upstream **DeckGlobalizer v2.1.1**\nby tinadu-ai (). Original\nthree-phase architecture (Visual Audit / Semantic Alignment /\nPage-by-Page Execution) credited and retained.\n\nFile v3.0.0:README.md\n\n# Deck Pipeline · CN→EN · McKinsey Polish · Layout Audit\n\n![version](https://img.shields.io/badge/version-0.1.0-blue)\n![license](https://img.shields.io/badge/license-MIT-green)\n![python](https://img.shields.io/badge/python-3.9%2B-blue)\n![status](https://img.shields.io/badge/status-beta-orange)\n\n> A production-grade Claude Code system for taking decks from raw Chinese draft to McKinsey-polished English — with full audit trail, layout integrity checks, and a swappable PROFILE block for project-specific defaults. Built on a 4-stage pipeline; also runs polish-only on any single-language deck.\n>\n> 一套生产级 Claude Code 系统，把 deck 从中文原稿做到麦肯锡级英文成稿——全程留痕、排版守护、可替换的项目级 PROFILE 块。基于 4 阶段流水线；也支持纯排版模式，处理任意单语言 deck。\n\n---\n\n## What it does / 做什么\n\nTakes a Chinese deck (and optionally an English draft) and runs it through a 4-stage pipeline. Or, if you give it a single-language deck and ask for layout-only cleanup, skips translation and runs polish stages alone.\n\nWorks for any kind of deck — sales, product, research, conference talks, internal reports, whatever.\n\n输入一份中文 deck（可选搭配一版英文草稿），经过 4 阶段流水线。或者，给它一份单语言 deck + 要求\"只做排版\"，就跳过翻译，只跑排版相关 stage。\n\n| Stage / 阶段 | What happens / 做什么 | Polish-only? |\n|---|---|---|\n| **1. Sense Pass** | Reverse-engineer design DNA (palette, fonts, size hierarchy) and surface candidate glossary terms.
反推设计 DNA（色板、字体、字号梯队），扒出候选术语。 | ✅ runs |\n| **2. McKinsey Translation** | CN → EN in McKinsey style from the first pass (top-down, parallel, strong verbs, no filler). Glossary applied inline; unclear terms asked on the spot. Runs **page-by-page** with a checkpoint after each slide.
中翻英直接出麦肯锡风格（top-down、平行结构、强动词、去填充词）。术语表 inline 应用；生僻词当场问。**逐页执行**，每页结束 checkpoint 等用户确认。 | ⏭️ skipped |\n| **3. Layout Audit** | Font pollution cleanup, overflow estimation, structural-anchor protection, reverse-sync from hand-edits.
字体污染清理、溢出估算、结构锚点保护、手动改动反向同步。 | ✅ runs |\n| **4. Handoff** | Three or four deliverables + a HANDOFF.md contract for the next session.
3–4 件交付物 + 给下一个 session 的 HANDOFF.md 接力契约。 | ✅ runs |\n\nYou do **not** need any external skill installed. The generic deck-globalization engine (3-phase visual audit / semantic alignment / page-by-page execution) is bundled inside.\n\n**不需要**安装任何外部 skill。通用 deck 全球化引擎（3 阶段：视觉审计 / 语义对齐 / 逐页执行）已内置。\n\n---\n\n## Install / 安装\n\n```bash\ngit clone https://github.com//deck-pipeline ~/.claude/skills/deck-pipeline\npip3 install python-pptx openpyxl pymupdf pyyaml\n```\n\nThe skill is loaded automatically by Claude Code on next start.\n\n下次 Claude Code 启动时会自动加载本 skill。\n\n---\n\n## Modes / 模式\n\n| Mode / 模式 | Trigger / 触发 | Stages run / 跑哪些 stage |\n|---|---|---|\n| **Full pipeline / 完整流水线** | CN deck (± EN draft) provided, translation requested
提供 CN deck（± EN 草稿），要翻译 | 1 → 2 → 3 → 4 |\n| **Polish-only / 纯排版** | Single-language deck, \"just polish / format only / skip translation\"
单语言 deck，说\"只做排版 / 跳过翻译 / format only\" | 1 → 3 → 4 |\n| **Reverse-sync only / 反向同步** | User hand-edited PPT after Excel was generated
用户在 Excel 生成后手动改了 PPT | 3.5 sub-routine |\n\n---\n\n## Outputs / 输出\n\n**Full pipeline / 完整流水线**（4 件）：\n\n| File / 文件 | Content / 内容 |\n|---|---|\n| `xxx-en-polished-[date].pptx` | English deck after McKinsey-style translation + glossary lock.
麦肯锡风格翻译 + 术语锁定后的英文 deck。 |\n| `xxx-final-[date].pptx` | Above + font/layout normalization, overflow fixes.
在前者基础上做完字体/布局规范化、溢出修复。 |\n| `xxx-bilingual-diff-[date].xlsx` | Row-by-row comparison: `page / kind / cn / en_original / en_optimized / notes`. Red-bold rows = key corrections.
逐行对照表，红粗 = 关键修正。 |\n| `HANDOFF.md` | Session contract: goal · tools · completed · unresolved · cautions · principles · constraints.
session 契约。 |\n\n**Polish-only / 纯排版**（3 件）：\n\n| File / 文件 | Content / 内容 |\n|---|---|\n| `xxx-final-[date].pptx` | After font/layout normalization, overflow fixes.
字体/布局规范化、溢出修复后。 |\n| `xxx-layout-changes-[date].xlsx` | What was changed.
改了什么。 |\n| `HANDOFF.md` | Same as above.
同上。 |\n\n---\n\n## Architecture / 架构\n\nThree layers, in order of stability / 三层，按稳定性排序：\n\n### L1 Tokens（stable, project-defined / 稳定，项目级定义）\n- **Palette / 色板** — ink, primary brand, soft fill, page background\n- **Font whitelist / 字体白名单** — title face (e.g. serif) + body face (e.g. sans-serif)\n- **Unit conversion / 单位换算** — 百万=M · 亿=100M · 十亿=1B · 百亿=10B · 千亿=100B · 万亿=1T\n- **Glossary / 术语表**（见下）\n\n### L2 Constants（operational defaults / 操作默认值）\n- **Size ladder / 字号梯队** — 22 / 14 / 10 / 8 / 6 / 4 pt\n- **Floors / 地板** — body ≥ 7pt · caption ≥ 6pt · source ≥ 4pt\n- **Compression step / 压缩步进** — `-0.1pt` discrete only / 仅 -0.1pt 离散迭代，不一刀切\n- **Line-spacing fallback / 行距备选档** — 1.25 → 1.15 before sub-floor compression\n- **Quote style / 引号风格** — McKinsey single quotes `'…'`\n- **Footer format / 页脚格式** — `Confidential · For Intended Recipients Only · {month} {year}` (middle dot, not em-dash)\n\n### L3 Sensed（run-time, per-deck / 运行时，按 deck 推算）\n- **Font-name suffix audit / 字体名后缀审计** — `font.name` must not contain `Bold` / `Regular` / `Italic` / `Light` suffixes; split into `name` + boolean attribute\n- **Structural-anchor detection / 结构锚点检测** — shapes consistent across ≥3 pages (footers, callout bands, watermarks) → per-page protection list\n- **Late-stage glossary re-scan / 末段术语复扫** — catches hand-edit regressions\n- **Glossary wavering detection / 术语摇摆检测** — same source term getting multiple translations across the deck\n\n---\n\n## Style baseline + extensible references / 风格基线 + 可扩展引用\n\n**McKinsey is the default baseline.** Title-as-takeaway, lead-with-so-what, parallel structure, strong action verbs, filler removal, sentence-case by default.\n\n**麦肯锡是默认基线**：标题即结论、so-what 前置、平行结构、强动词、去填充词、默认小写。\n\n**On top of McKinsey, you can upload additional reference samples** — any PDF or .pptx whose writing style you want to emulate (a colleague's well-written report, an industry whitepaper, your own prior work, etc.). The skill distills each via `scripts/style_distill.py` and layers its rules on top of the McKinsey base.\n\n**在麦肯锡之上，可以上传其他参考样本**——任何你想模仿其文风的 PDF 或 pptx（同事写得好的报告、行业白皮书、自己之前的作品等）。skill 用 `scripts/style_distill.py` 抽每份样本的风格指纹，在麦肯锡基线之上叠加。\n\nDistillation extracts / 蒸馏出来的内容：\n- Cadence: avg sentence/paragraph length, p90 length / 节奏：句长、段长、p90\n- Vocab: signature phrases, top action verbs, filler-word patterns / 词汇：标志短语、高频动词、填充词模式\n- Structure: bullet pattern, parallel-structure score / 结构：bullet 模式、平行度\n- Tone: first-person ratio, hedge ratio, certainty ratio / 调性：第一人称比例、保守语 / 笃定语比例\n\nAdd references in PROFILE / 在 PROFILE 加 references：\n\n```yaml\nstyle_references:\n - path: \"/path/to/sample.pdf\"\n weight: 0.7 # 0.0–1.0; how strongly to bias toward this sample\n```\n\nConflicts between references → most recent entry wins; user asked at first conflict.\n\nReferences 冲突 → 最新一条优先；首次冲突时问用户。\n\n---\n\n## Page-by-page execution / 逐页执行\n\nStage 2 is **not** all-at-once. After overall confirmation, the skill iterates slides 1 → N. For each slide:\n\nStage 2 **不是**一次性全跑。整体确认后逐页跑：\n\n1. Apply edits to that slide's paragraphs / 应用该页的 edits\n2. Append rows to the Excel immediately / 立刻把该页的对照行追加到 Excel\n3. Print \"P{n} done — N changes applied. Continue?\" / 打印 \"P{n} 完成，N 处改动，继续？\"\n4. Wait for user confirmation before moving to P{n+1} / 等用户确认再进下一页\n\nYou can interject \"back to P{n-1}\", \"stop here\", or \"redo P{n}\" between pages.\n\n页间可以说\"回到 P{n-1}\"、\"停在这里\"、\"P{n} 重做\"。\n\nWhy / 为什么：\n- Mid-stream review and steering / 中途可审查、可调整\n- Bad assumptions don't propagate silently / 错误假设不会静默扩散\n- Excel grows incrementally → survives interruption / Excel 增量增长，中断也能续上\n- Token-efficient: only current slide in active scratchpad / 省 token：当前页才在活跃上下文\n\n---\n\n## Glossary categories / 术语表分类（extensible / 可扩展）\n\n```\nLocked → user-confirmed, never re-asked / 用户确认过，不再问\nPending → asked but not yet confirmed / 问过但未拍板，跨 session 接力\nRejected rewrites → user vetoed; never propose again / 被驳回过，永久回避\nSession-added → added mid-session; promoted to Locked at handoff / 本次新加，handoff 时升 Locked\nPer-row override → \"ignore glossary for this row\" / 单行豁免\n```\n\nReplace the example entries in `SKILL.md` `PROFILE.glossary.locked` with your project's terms (proper nouns, brand names, product names, industry phrases, etc.).\n\n把 `SKILL.md` 中 `PROFILE.glossary.locked` 的示例条目替换成你项目的术语（专有名词、品牌名、产品名、行业说法等）。\n\n---\n\n## Magnitude trap / 量级陷阱\n\nThe Chinese 亿 is **100 million**, NOT \"billion\". / 中文「亿」= 1 亿 = 100M，**不是** \"billion\"。\n\n| CN | EN |\n|---|---|\n| 一百万 / 百万 | 1 million |\n| 一千万 / 千万 | 10 million |\n| 一亿 / 1亿 | 100 million |\n| 十亿 | 1 billion |\n| 百亿 | 10 billion |\n| 千亿 | 100 billion |\n| 万亿 | 1 trillion |\n\nAppend your currency suffix (`$`, `RMB`, `€`, etc.) in PROFILE per project.\n\n---\n\n## Interaction model / 交互模型\n\nThe skill is **bidirectional** — it stops and asks at specific checkpoints rather than guessing.\n\n本 skill **双向交互**——在特定检查点停下来问，不瞎猜。\n\nYou will be asked when / 以下情况会问你：\n\n1. A source term has no glossary entry and Claude is unsure
遇到未收录术语且拿不准\n2. A McKinsey rewrite candidate looks ambiguous (only when unsure — not every line)
麦肯锡改写候选有歧义（仅在拿不准时问，不是每行都问）\n3. A shape's font hits the floor and overflow remains
某 shape 字号撞地板但仍溢出\n4. A companion file (`.pptx` / `.xlsx`) is open in another app (lock file present)
配套文件被其他 app 打开（lock 文件存在）\n5. The overflow estimator reports > 10 HIGH-risk shapes (defer to user-side rendering)
估算器报 > 10 个 HIGH 风险 shape（让用户外部渲染复核）\n\nYou will **not** be asked twice for the same thing in one session — once confirmed, it enters Locked.\n\n同一件事**不会问两次**——确认过的进 Locked。\n\n---\n\n## Operational rules / 操作规则\n\n### File-write discipline / 文件写入纪律\n- Overwrite the original by default / 默认覆盖原文件\n- Scan for `~$xxx` lock file before writing / 写前扫 `~$xxx` lock 文件\n- If lock exists → **stop and ask user to save + close first** / 存在 → **停 + 让用户先保存并关闭**\n\n### Excel companion guard-rails / Excel 配套文件三护栏\n1. **Pre-write check / 写前校验** — verify header has all 6 columns / 校验 header 6 列齐全\n2. **Post-write readback / 写后回读** — immediately reload and confirm columns survived / 立即重读，确认列数没丢\n3. **Reverse sync / 反向同步** — bidirectional diff (ordinal first, fuzzy fallback) updates Excel from a hand-edited PPT / 双向 diff（先 ordinal，再 fuzzy 兜底），把 PPT 的修改同步回 Excel\n\n### Font compression discipline / 压字号纪律\n- Detect per-page protection list (structural anchors + user-specified) / 先取 per-page 保护列表（结构锚点 + 用户指定）\n- Iterate `-0.1pt` until no overflow / 按 -0.1pt 迭代到无溢出\n- If floor is hit / 撞地板：\n 1. Geometric widening (eat margin) / 横向扩框（吃 margin）\n 2. Line-spacing 1.25 → 1.15 / 行距压缩\n 3. Still overflowing → **escalate to user** / 仍溢出 → **上报用户**，列出问题 shape\n\n### CN-alignment configuration / CN 对齐配置\nConfigure `PROFILE.cn_en_slide_offset` (or pass `--cn-offset ` to `excel_sync.py`) for decks where CN and EN have been restructured.\n\n如果 CN/EN 两版有结构性差异，配置 `PROFILE.cn_en_slide_offset`（或给 `excel_sync.py` 传 `--cn-offset `）。\n\nDefault is 1:1 alignment.\n\n默认 1:1 对齐。\n\n---\n\n## Overflow estimator / 溢出估算器\n\nStatic analyzer flagging shapes whose text likely overflows. Three severity bands:\n\n静态分析，三档严重度：\n\n- **HIGH** — `ratio > 1.5`（very likely real / 极可能为真）\n- **MED** — `ratio 1.15 – 1.5`（might be real / 可能为真）\n- **LOW** — `ratio 1.0 – 1.15`（probably fine / 大概率没事）\n\nOnly HIGH is surfaced by default. / 默认只报 HIGH。\n\n### Accuracy details\n- Honors `auto_size`（NONE / SHAPE_TO_FIT_TEXT / TEXT_TO_FIT_SHAPE）\n- Reads actual margins (no defaults)\n- Line-height multiplier `1.15`\n- Per-character width by class (narrow / wide / digit / upper / space)\n- Greedy word-wrap simulation\n\n### When the estimator is still wrong / 估算器仍可能错\n\nStatic analysis cannot render. If > 10 HIGH-risk shapes are reported:\n\n静态分析不能渲染。HIGH > 10 个时：\n\n> \"Render the deck to PDF or PNGs externally (Keynote `File → Export`, or PowerPoint `Save as PDF`) and tell me which page numbers look problematic. I'll fix those targeted pages.\"\n>\n> 「请用 Keynote 导出或 PPT 另存 PDF，告诉我哪几页有问题，我针对性修。」\n\nBy design — dumping 30+ rendered pages into a single session causes context overload.\n\n这是刻意设计——把 30+ 渲染页一次性灌进 session 会过载。\n\n---\n\n## Known limitations / 已知限制\n\n1. Overflow estimator is a **hint, not a verdict** — final visual check requires external rendering.
溢出估算器是 hint 不是判决，最终视觉确认需外部渲染。\n2. `python-pptx` cannot render slides.
`python-pptx` 不能渲染。\n3. CN auto-alignment can drift on heavily restructured pages — configure overrides for known cases.
CN 自动对齐对重排页可能漂移，已知情况配 overrides。\n4. The skill assumes the CN source is semantic ground truth — typos in CN propagate.
把 CN 当语义基线，CN 自带的错也会传过去。\n5. File-lock collisions silently corrupt output. Always close before letting the skill write.
文件锁冲突会静默破坏输出，让 skill 写之前一定要关。\n\n---\n\n## License / 许可\n\nMIT. See `LICENSE`. / MIT 协议，见 `LICENSE`。\n\n## Contributing / 贡献\n\nSee `CONTRIBUTING.md`. / 见 `CONTRIBUTING.md`。\n\n## Credits / 致谢\n\nGeneric deck-globalization engine derived from upstream **DeckGlobalizer v2.1.1** by tinadu-ai (). Original three-phase architecture (Visual Audit / Semantic Alignment / Page-by-Page Execution) credited and retained.\n\n通用 deck 全球化引擎来自上游 **DeckGlobalizer v2.1.1**（作者 tinadu-ai，）。原三阶段架构（视觉审计 / 语义对齐 / 逐页执行）credit 保留。\n\nFile v3.0.0:_meta.json\n\n{\n \"ownerId\": \"kn7c4vnnxm0hrg85thcqxz9h3181tzkd\",\n \"slug\": \"deck-pipeline\",\n \"version\": \"3.0.0\",\n \"publishedAt\": 1778603607011\n}\n\nFile v3.0.0:CHANGELOG.md\n\n# Changelog\n\nAll notable changes are documented here.\nFormat follows [Keep a Changelog](https://keepachangelog.com/).\n\n## [Unreleased]\n\n## [0.1.0] — 2026-05-12\n\n### Added\n- Initial public release.\n- 4-stage pipeline: Sense Pass → McKinsey Translation → Layout Audit → Handoff.\n- Polish-only mode for single-language decks (skips translation).\n- Swappable `PROFILE` block (palette, fonts, glossary, style rules, slide-offset config).\n- 10 helper scripts:\n - `sense_pass.py` — reverse-engineer design DNA (palette, fonts, size hierarchy).\n - `extract.py` — paragraph-level text extraction.\n - `apply.py` — apply EN edits + write comparison Excel.\n - `layout_audit.py` — font pollution cleanup + font-name suffix audit.\n - `overflow_recheck.py` — overflow estimator with HIGH/MED/LOW tiers.\n - `anchor_detect.py` — cross-page structural-anchor detection.\n - `glossary_audit.py` — late-stage glossary re-scan + wavering detection.\n - `excel_sync.py` — bidirectional PPT ↔ Excel sync, configurable CN↔EN slide alignment.\n - `handoff.py` — HANDOFF.md template writer.\n - `style_distill.py` — distill style fingerprint from a reference PDF/.pptx.\n- Page-by-page execution with per-slide user checkpoint.\n- Extensible style references (McKinsey baseline; users can layer additional reference samples).\n- Bidirectional PPT ↔ Excel sync (ordinal-position + fuzzy fallback).\n- Overflow estimator improvements: honors `auto_size`, reads real margins, per-character width by class, greedy word-wrap simulation.\n- Font-name suffix audit (`Bold`/`Regular`/`Italic`/`Light` not allowed in `font.name`).\n- Structural-anchor cross-page detection (auto per-page compression protection list).\n- Discrete `-0.1pt` font compression step; line-spacing fallback (1.25 → 1.15).\n- Companion-file lock detection (`~$xxx`).\n- Excel three guard-rails (pre-write check, post-write readback, reverse sync).\n- HANDOFF.md as a session-end deliverable.\n- Bilingual README (English + 中文).\n\n### Sources / credits\n- The generic deck-globalization engine is derived from upstream\n **DeckGlobalizer v2.1.1** by tinadu-ai\n ().\n Original three-phase architecture (Visual Audit / Semantic Alignment /\n Page-by-Page Execution) credited and retained.\n\nFile v3.0.0:CONTRIBUTING.md\n\n# Contributing\n\nThanks for your interest. This is a small, opinionated tool — contributions are welcome but please keep the surface area minimal.\n\n## Reporting issues\n\nOpen a GitHub issue with:\n- What you ran (command + relevant CLI flags)\n- What you expected vs what happened\n- A minimal reproduction (a sanitized 1–3 slide `.pptx` is ideal)\n- Your environment (Python version, OS, `python-pptx` version)\n\nDo **not** attach decks containing confidential content.\n\n## Pull requests\n\n1. Fork → branch from `main` → PR back into `main`.\n2. Keep PRs focused. One concern per PR. Split if needed.\n3. Touch the smallest possible surface.\n4. Update `CHANGELOG.md` under `[Unreleased]` with a one-line summary.\n5. If you change script behavior, update the corresponding section in `SKILL.md` and `README.md`.\n6. New scripts must:\n - Be runnable standalone via `python3