gorden-ppt-gen

name: gorden-ppt-gen description: >- Generate an "image-format PPT" from a topic/content — one high-density, complex-layout, style-consistent rendered image per slide via an image-gen model (default tech-business high-density infographic; any style supported), then compose into a full-bleed-image .pptx. 주제/내용을 이미지 PPT로 생성, AI 출도 슬라이드, 이미지 버전 PPT 제작. Use when "이미지 PPT 만들어줘", "PPT 이미지로 생성", "AI 슬라이드 출력", "make image PPT", "generate slide deck images". Produces ONLY image-PPT (one .png per slide + composed .pptx); for an editable .pptx pipe into gorden-img2pptx. Do NOT use for editable/template decks (use ppt-template-engine), Jobs-style keynote decks (use keynote-architect), or one-shot topic→editable (use gorden-super-ppt). Ported from @Gorden Sun (github.com/GordenSun/GordenSuperPPTSkills), adapted for Claude Code.

gorden-ppt-gen — Generate Image-format PPT

Turn a topic/content into an image-PPT: design → generate → compose. Each page is one model-generated image. Must use raster image generation — never SVG/HTML/Canvas/code drawing.

content ─▶ A2 outline.json ─▶ A3 per-page prompts ─▶ A4 image-gen ─▶ A5 compose image .pptx

Read references/image-prompt-guide.md for the full prompt system, complexity checklist, style library, chart catalog, and verbatim rules. Read references/runtime-notes.md for the Claude Code image backend.

Backend (Claude Code adaptation)

Codex imagegen is replaced by:

• Primary: python3 ../gpt-image-2/scripts/generate_image.py --prompt "<full page prompt>" --size 1536x1024 --quality high --output slides/NN.png --style none (needs OPENAI_API_KEY). 16:9 not native → generate with safe margins then center-crop (see runtime-notes §5).
• Fallback: mcp__hf__gr1_z_image_turbo_generate (main agent only).
• A4 must write imagegen-manifest.json per page (slide, prompt_file, generated_source, copied_to, backend). No manifest or any page missing generated_source → skill fails. If no backend available, stop and explain — never fall back to code-drawn PNGs.

imagegen hard gate

"Generate image" means call an image model, NOT draw a PNG with code. Banned for whole-page slides: Python/PIL, SVG, HTML, Canvas, matplotlib, PowerPoint shapes, screenshot rendering, or building an editable PPTX then exporting to PNG. Code may ONLY: write outline/prompts/manifest/deck JSON, copy generated images into slides/, crop/scale a whole generated image for aspect ratio (no text patching), compose the image PPTX, QA.

Global laws

1. 🔴 No placeholder images — output finished slide images containing ALL real text verbatim, not empty templates. The prompt's page-text must be full real text. lorem/empty boxes/text-less cards = fail → fix prompt, regenerate.
2. Zero fabrication — use the user's data/values/proper nouns exactly. Thin data → make content thick (real structured breakdown) + large-type/center-focus layout; never invent.
3. Language follows user — Q&A in user language; slide text in content language.
4. Avoid large pure-green areas (downstream gorden-img2pptx chroma-keys green).
5. Clean footer — no page numbers/logo/watermark unless asked.

Default style & complexity

Default = very high information density, complex layout, modular grid, a different advanced framework per page (house/Möbius/Bento/concentric-radar/multi-layer architecture/fishbone/funnel/3D stairs). Simple = fail.

• ⭐ Content first (most common pitfall): "layout too simple" almost always = each page too thin. Prep thick content first: each page ≥4 parallel modules, each = title + 2-3 points + 1 huge metric, plus a lead-in (keyword highlight) + bottom banner (≈20+ info points/page). See guide §0.
• No two pages alike + unified palette; extract palette from a reference if given.
• Luxe ≠ gaudy — clarity first; "clear-tech" (light bg + fine grid + flat modules + restrained emphasis + high density) is the safe default.
• Layout references in reference-layouts/: pick a structurally similar one, write its layout skeleton into the prompt (or pass as reference) — learn layout, never copy its text/brand.
• Aspect ratio follows user; default 16:9.

Workflow (check each box)

- [ ] A1 Confirm: style / audience / page count / language (skip only if user says "just generate")
- [ ] A2 outline.json: deconstruct content → assign a non-repeating complex framework per page, unified palette, thick detailed_content (guide §7/§1.7)
- [ ] A3 Per-page prompts: land outline into self-contained prompts (with ALL real text verbatim) → prompts/NN-*.md
- [ ] A4 Generate: per page call the backend (full real text, no placeholder) → copy into slides/ → write imagegen-manifest.json
- [ ] A5 Compose: compose_pptx.py builds a full-bleed image .pptx from deck.json

• Pages: user count if given, else 11-20 (cover + closing always).
• A3 self-contained: hardcode colors(hex)/layout/all-text/chart-form/font-style for reproducibility + single-page reruns.
• A4: all prompts/NN-*.md ready before starting; one real generate call per page; only retry the failed page; keep source originals.
• A5: write deck.json (each slide background → its PNG), then:

python3 scripts/compose_pptx.py deck.json out/<topic>-image-deck.pptx

Output / scripts

<topic-slug>/{outline.json, imagegen-manifest.json, prompts/NN-*.md, slides/NN-*.png, deck.json, out/<topic>-image-deck.pptx}

pip3 install python-pptx pillow. Next: editable pptx → gorden-img2pptx; one-shot → gorden-super-ppt.