deck

name

deck

description

General-purpose presentation/deck slide generation as gpt-image renders via the codex CLI — any brand, any theme, KO / EN / mixed language. Theme-intake from user assets with a neutral zero-config fallback, language-aware quality gates, numbers whitelist, and deterministic Pillow post-processing. Triggers - "deck", "slides", "presentation images", "pitch deck", "make slides for", /deck. For the IntuneLabs IR house style specifically, use /ppt instead.

deck — brand-agnostic slide generation (codex CLI, gpt-image)

General-purpose sibling of ppt. ppt is the mature, frozen IntuneLabs-KO preset; this skill is the canonical home of the shared mechanics (whitelist, gate, routing, compositing). If a mechanic improves, change it HERE first and backport deliberately — two drifting copies is the known failure mode (fresh-eyes 2026-07-08).

When to use

• Anyone (not just IntuneLabs) asks for presentation/IR/pitch slides as rendered images.
• The deck is English, Korean, or mixed.
• The user has their own brand (logo/palette/reference slide) — or no brand at all.

Interface: asset-first, default-heavy

Ask the user for AT MOST four things; default everything else structurally:

1. Language — auto-detect from their content; ask only when ambiguous. Values: ko / en / mixed.
2. Brand assets — logo file, palette, an approved reference slide, real screenshots. None → neutral theme.
3. Deck intent — audience, slide count, source content/facts.
4. Strict strings & numbers — names, dates, figures, quotes that must appear verbatim.

Never run a long questionnaire. Extract, don't interrogate.

Step 0 — Theme contract (before any render)

Fill templates/theme-contract-template.yaml and show it to the user for a quick OK. Sources, in order of preference:

1. Reference slide provided → extract palette hexes, wordmark position, title grammar, density by inspection. Best signal; skip most questions.
2. Logo + palette only → adopt the neutral layout grammar with their colors/wordmark.
3. Nothing → the built-in neutral theme: off-white bg #F5F6F8, white panels with slate-navy header bars #3D5A73, charcoal text #2D3436, one accent #C0392B used only for the single emphasis phrase per slide, no wordmark required, page number bottom-center.

Intake validations (each one is a known silent breaker):

• Contrast check — panel-header text vs bar color, body vs bg ≥ WCAG-ish 4.5:1. Dark/low-contrast palettes also break the --dark-bbox phone detection in compositing → note screenshot_policy.patch_mode: quad in the contract when bg is dark.
• Wordmark polarity — white-on-transparent logos vanish on light bg; check the alpha channel (glyphs may live ONLY in alpha; never convert('L') blindly) and pick/request the dark variant.
• Font reality — the image model cannot load brand fonts. Record a nearest describable family mapping (e.g. "Pretendard-like clean sans", "Georgia-like serif") in typography, and the REAL font file only for Pillow micro-text patches.
• Layout grammar is brand identity — the two-line arrow title / takeaway bar grammar is a preset, not a universal. If the reference slide shows a different grammar, encode THEIRS in title_grammar / layout_slots; don't ship someone a recolored IntuneLabs deck.

Escape hatch: if the brand requires pixel-exact identity (strict brand guidelines, legal marks), genAI rendering is the wrong tool — offer deterministic HTML→PNG instead.

Language handling (the "full bilingual" contract)

mixed is NOT a vague third mode — every critical string carries a language tag; each string gets its own language's gate. Structural defaults per language:

concern	ko	en	mixed
prototype label	시제품	PROTOTYPE	both, per surface language
spell gate	Hangul char-by-char + KO killer list	word-by-word + EN killer list + casing check	per-string by tag
patch font	AppleSDGothicNeo.ttc	Helvetica/SF (or contract font)	per-string
footnote prefix	1) …	1) …	shared

KO killer list (seed; grow per project): 코드로≠크드로 · 웹훅≠웹혹 · 플러그인≠풀러그인 · 통계청≠통례청 · 캡처≠캡쳐 · 점수화≠잠수화 · 언제나≠연제나 · 친화≠천화. EN failure modes & killer list (seed): pseudo-words and dropped/doubled letters (Recommendation, Infrastructure, Institution, Governance, Efficiency, Acquisition, Comparative, Loneliness), casing drift (Title Case↔ALL CAPS↔sentence case must match the anchor), fake ligatures/kerning artifacts, confusables l/1 · e/o · rn/m, apostrophes (it's/its), broken hyphenation, invented abbreviations. Put CRITICAL SPELLING callouts in the prompt for any killer word that appears; for EN also pin the exact casing: render exactly "Global Launchpad" — capital G, capital L, no hyphen. Mixed-only gate items: (a) no script swap — a string tagged KO must not come back in English or vice versa; (b) terminology-pair consistency — a KO term and its EN partner (e.g. 시제품/PROTOTYPE) stay paired identically across all slides; (c) font harmony — Hangul and Latin must read as one family, call it out in the style paragraph.

Generation mechanics (inherited, proven — see ppt for war stories)

Codex CLI on the user's ChatGPT subscription (never export/print OPENAI_API_KEY):

codex exec --skip-git-repo-check -i "<anchor.png>" -i "<wordmark.png>" [-i "<shot.png>"] < prompt.md

• Prompt via stdin; -i order matters (prompt says "FIRST/SECOND/THIRD attached image").
• Output: ~/.codex/generated_images/<session>/ig_*.png at 1672×941. codex LIES about paths — accept only paths under ~/.codex/generated_images/, fallback find -newermt.
• ~5–10 min/render; run_in_background: true, timeout ≥ 1200000; ≤4 concurrent; >8 slides → Workflow (author → generate → gate, batches of 4).
• Prompt skeleton: templates/slide-prompt-template.md (parameterized by the theme contract — fixed frame, title grammar, density rule, spelling callouts, numbers whitelist).
• Anchor discipline: render slide 1, get user approval, then attach it as ref 1 to every later render. The anchor IS the theme contract made visible.

The gate (run on EVERY render, viewed with Read)

1. Spelling/casing, language-aware — per-string by language tag (see above), 0 defects.
2. Numbers whitelist — every visible digit ⊆ the prompt's exhaustive allowed list.
3. Screenshot preservation — real screenshots never redrawn; garbled → composite queue.
4. Layout/density — matches the theme contract's grammar and density rule; no stray top numerals.
5. Honesty guardrails — no unverified claims; prototypes labeled (시제품/PROTOTYPE); in-progress infra described as in-progress; no fake facility/team photos.
6. Anchor consistency — palette/typography/frame vs the approved anchor.
7. (mixed only) Script-swap & terminology-pair check. Retry ≤4 with targeted feedback; same fix fails twice → stop, best render + deterministic patch.

Deterministic post-processing

Scripts are shared with ppt via the scripts/ symlink (single canonical copy):

• composite_screenshot.py — real pixels into rendered frames (--quad / --dark-bbox / --flood). Dark-theme caveat: --dark-bbox assumes dark phone on light bg; on dark themes use --quad with crop-verify.
• contact_sheet.py — N-up review montage.
• Micro text / verbatim-critical strings (dates, names, quotes, legal lines) are ALWAYS typeset deterministically with the per-language patch font — genAI garbles them plausibly.
• Frame normalization at merge: stamp wordmark + page numeral at fixed coords on all slides, then resize 1672×941 → 1920×1080.

Loud/silent routing (unchanged, load-bearing)

• genAI: layout, diagrams, display-size text — failures are loud, the gate sees them.
• Pillow: micro text, verbatim-critical strings, anything that failed the gate twice.
• Real faces: only with the person's real photo attached as ref, rendered large, plus a same-person gate item; drift or small faces → composite real pixels. Tell the user which faces are genAI-passed.

Success criteria

• Theme contract user-approved before slide 2; anchor approved before the fleet
• Every render passed the gate or has a merge-queue entry
• All screenshots are real pixels; verbatim strings typeset, not generated
• Numbers ⊆ whitelists; language gate clean per string tag
• Contact sheet delivered; user approves before final assembly

File layout

~/.claude/skills/deck/
├── SKILL.md
├── templates/
│   ├── theme-contract-template.yaml
│   └── slide-prompt-template.md
└── scripts -> ../ppt/scripts        # single canonical copy, do not fork