Deck Pipeline

A 4-stage, audit-trailed Chinese→English deck globalization system with a swappable project profile.

This skill bundles the generic deck-globalization engine (originally upstream DeckGlobalizer v2.1.1) and an editable PROFILE block (palette, fonts, glossary, style preferences). The two are separated by section so the profile can be swapped per project / brand without touching the engine.

For a marketing-style overview, see README.md in this directory. For implementation, see scripts/ and the per-stage runbooks below.

0. Modes & activation

Detect the mode in the first turn. If ambiguous, ask one yes/no question ("This deck is already in EN — should I just polish layout, or also rewrite McKinsey-style?"). Do not guess silently.

1. PROFILE block — defaults (swappable)

Edit this block to retarget the skill for your project / brand. Everything below this block is profile-agnostic.

PROFILE:
  # ---- L1 Tokens ----
  palette:
    # Replace with your brand colors.
    ink:          "#1A1A1A"
    primary:      "#000000"   # accent / brand primary
    soft:         "#FFFFFF"   # soft fill behind banners
    page_bg:      "#FFFFFF"
  fonts:
    # Choose a serif title face + a sans-serif body face for best contrast.
    title:        "Georgia"
    body:         "Verdana"
    title_bold:   true
  unit_table:
    # Chinese number magnitudes → English. 亿 is 100M, NOT "billion".
    "百万":       "M"
    "千万":       "10M"
    "亿":         "100M"
    "十亿":       "1B"
    "百亿":       "10B"
    "千亿":       "100B"
    "万亿":       "1T"
    # Currency suffix is left to the user — append "$" / "RMB" / "€" as appropriate.

  # ---- L2 Constants ----
  size_ladder:           [22, 14, 10, 8, 6, 4]   # H1, H2, body, caption, footnote, source
  floors:
    body:     7
    caption:  6
    source:   4
  compression_step:      0.1     # discrete -0.1pt iterations only
  line_height_default:   1.25
  line_height_fallback:  1.15    # used before sub-floor compression
  quote_style:           "single"  # 'McKinsey' single quotes
  footer_format:         "Confidential · For Intended Recipients Only · {month} {year}"
  separator_in_footer:   "·"     # middle dot, NOT em-dash

  # ---- L1 Glossary (extensible) ----
  # Replace the example entries below with your project's locked terms.
  # Categories are illustrative; you can rename / add / remove.
  glossary:
    locked:
      people_orgs:
        # "<source term>": "<canonical translation>"
        # e.g. "John Smith": "John Smith"
        # e.g. "Acme Capital": "Acme Capital"
        {}
      business_terms:
        # Common Chinese business-deck idioms with industry-standard
        # English mappings. Edit / extend as needed.
        "流水":     "gross revenue"
        "私域":     "owned audience"
        "出海":     "global expansion"
      domain_specific:
        # Project / industry / domain terms.
        # "<source term>": "<canonical translation>"
        {}
    rejected_rewrites:
      # Entries the user vetoed during prior sessions.
      # Format: { source: "...", proposed: "...", reason: "..." }
      []
    pending: []
    session_added: []

  # ---- Style rules ----
  # McKinsey is the default baseline. Additional style references can be
  # uploaded and distilled via scripts/style_distill.py; their rules layer
  # ON TOP of the McKinsey base.
  style_baseline: "mckinsey"
  mckinsey:
    title_is_takeaway:      true   # title = the so-what, not the topic
    lead_with_so_what:      true
    parallel_structure:     true   # bullets share tense, opening part-of-speech
    strong_action_verbs:    true   # cut "is/has", prefer concrete verb
    cut_filler:
      - "in order to → to"
      - "a number of → many"
      - "due to the fact that → because"
      - "at this point in time → now"
    case:                   "sentence"   # lowercase unless proper noun or locked term
    em_dash_policy:         "use em-dash for parentheticals; use · (middle dot) in lists/footers"
  style_references:
    # Each entry is a PDF / .pptx reference. style_distill.py reads it and
    # emits rules (cadence, signature phrases, paragraph length, tone) that
    # layer on top of the McKinsey base. Conflicts: more recent entry wins;
    # user is asked at first conflict.
    # Example:
    # - path: "/path/to/sample.pdf"
    #   weight: 0.7
    []

  # ---- Structural anchor heuristics ----
  anchor_detection:
    min_pages: 3              # appears on ≥3 slides
    match_on:                 # signature components
      - position_xy
      - fill_color
      - font_size_class
    auto_protect: true

  # ---- Overflow estimator ----
  overflow:
    severity:
      high: 1.5
      med:  1.15
      low:  1.0
    surface_only: "high"      # surface MED/LOW only when explicitly asked
    defer_to_user_threshold: 10   # if HIGH > 10 → ask user to render externally

  # ---- CN ↔ EN slide alignment ----
  # Default is 1:1 (EN slide N maps to CN slide N).
  # Set overrides only when the two decks have been restructured.
  # Pass this config to excel_sync.py via `--cn-offset <yaml>`.
  cn_en_slide_offset:
    default: 0           # offset added to EN slide number (0 = 1:1)
    overrides: {}        # e.g. {"9-26": -1, "20": null}
                         # int = relative offset; null = no CN counterpart

Profile-agnostic note: all sections below treat PROFILE as an opaque dict. Do not hardcode project-specific values anywhere outside the PROFILE block.

2. Pipeline stages

Each stage has: inputs · what it does · outputs · stop-and-ask conditions.

Stage 1 — Sense Pass

Inputs: one or two .pptx paths (CN, optional EN draft) What it does:

1. Run scripts/sense_pass.py to extract:
   • palette (top fill colors)
   • font usage histogram
   • size distribution
   • title-zone shapes (top ≤ 600K EMU)
   • layout heuristics
2. Cross-check sensed values against PROFILE.palette / PROFILE.fonts. If a sensed font is NOT in the whitelist AND NOT in SKIP_POLLUTION, record it as font pollution.
3. Surface candidate glossary entries: any CN noun phrase that appears ≥2 times and isn't already in glossary.locked.

Outputs:

• Style_Manifest.md (in-memory; not written to disk unless requested)
• pollution_report (slide → font → count)
• candidate_glossary (term → count → sample context)

Stop-and-ask:

• Candidate glossary surfaces a term Claude can't confidently translate → ask user, write answer to glossary.session_added.
• Sensed primary palette color differs from PROFILE.palette.primary → ask whether to update profile or keep existing.

Stage 2 — McKinsey Translation (skipped in polish-only mode)

Inputs: Stage 1 outputs + the CN deck (and optional EN draft for diff context) + any uploaded style_references.

Style layering: McKinsey base rules (PROFILE.mckinsey) apply first. If PROFILE.style_references is non-empty, run scripts/style_distill.py on each reference before translation begins; the distilled rules (cadence, signature phrases, paragraph length, tone) layer on top. More recent entry wins on conflict; ask user at the first conflict.

Page-by-page execution (hard requirement):

1. Overall confirmation first — after Stage 1, show the user the planned per-page edit count + sample of style rules in effect; wait for "go".
2. Then loop slides 1 → N, one at a time:
   • Collect paragraph-level CN text on this slide via scripts/extract.py.
   • For each paragraph, produce EN per the layered style rules:
     • lowercase by default; title =