_{<b>🌐 English</b> · <a href="README.zh.md">中文</a>}

huashu-slide-codex

Codex-only AI visual material production skill.
Say one sentence — get a finished PPT, WeChat cover, or Bilibili/YouTube thumbnail.

Why Codex-only?

Most "AI PPT" skills wrap Gemini / OpenAI Image / Nano Banana API calls — every slide costs money. Codex has built-in image_gen: it's already paid for via your Codex subscription. The whole point of this skill is to use that built-in capability instead of burning a second image API.

If you run this skill in Claude Code or Cursor, the built-in image_gen doesn't exist there, and the skill won't work as intended. Use huashu-design (HTML-native) or huashu-wechat-image (Gemini-backed) for those runtimes.

What it does

Four delivery paths, all powered by Codex's image_gen:

Quick start

In a Codex session:

Use huashu-slide-codex to make a 10-page deck about [your topic].

That's it. The skill will:

1. Ask 1-3 clarifying questions (audience, tone, format)
2. Recommend 3 differentiated design philosophies (Bloomberg / Field Notes / Pentagram etc.)
3. Run a brand asset protocol if your topic involves a specific brand (downloads logo, product shots, UI screenshots → writes a project-local brand-spec.md)
4. Generate slides via image_gen, copy into your project, assemble PPTX + HTML

For a single cover image:

Use huashu-slide-codex to make a WeChat cover for my article about [topic].

The skill defaults to 3 differentiated versions for important single images, so you have real choices instead of "the AI's best guess."

Install

cd ~/.codex/skills/
git clone https://github.com/alchaincyf/huashu-slide-codex.git

Or wherever your Codex skill directory lives.

Optional: image hosting

If you want to publish results to WeChat or other platforms that need permanent URLs, set up ImgBB:

cp .env.example .env
# Edit .env and add your free ImgBB key from https://api.imgbb.com

The bundled scripts/upload_image.py uses Python stdlib only — no pip install needed.

Core mechanics

🔴 Default path lock

In Codex environments, the skill always defaults to Path 1 (AI Image PPT). The two — and only two — triggers for switching to Path 3 (HTML/editable) are:

1. User explicitly says "I want editable PPT" / "don't want image PPT"
2. image_gen actually fails to invoke (≥3 times)

The skill includes an explicit list of "forbidden self-justification" patterns to stop the agent from rationalizing away the default. This was added after a real-world failure where Codex talked itself into HTML mode because "the content needs precise version numbers."

Page types & density tiers

Not every slide should have the same density. The skill enforces 4 page types:

The skill explicitly warns the agent that the ceiling is not a target — "the sparse content page is more professional than the stuffed one."

Brand Asset Protocol

If your topic involves a specific brand (Anthropic, Linear, your own company), the skill runs a 5-step protocol:

1. Ask user for what assets they have (logo / product shots / UI screenshots / brand guidelines)
2. Search official channels (<brand>.com/brand, /press-kit, press kits, App Store)
3. Download via curl
4. Verify + extract colors (grep hex codes from inline CSS)
5. Write project-local brand-spec.md + mandatory user checkpoint before generating any slides

The protocol exists because cover images / decks without real brand assets become "generic AI tech aesthetic." 30 minutes of asset gathering saves 2 hours of rework.

Spread Frame + Mascot Continuity

For information-manual style decks (Field Notes × Anthropic, Penguin Books, Bloomberg Businessweek), the skill recommends:

• Spread Frame: A consistent 3-element layout across all content pages — top brand bar + free body + bottom conclusion strip. Body layout varies, frame stays.
• Mascot Continuity: If your brand has a mascot (a shrimp, a pixel character, an IP figure), have the mascot appear across content pages in different poses — turns a deck from "8 isolated infographics" into "a connected story."

Both patterns are validated by real-world delivery (OpenClaw Orange Paper PPT, 2026-05-23).

Personal-brand asset injection (swappable)

The skill ships with three example pixel-style assets in assets/personal-brand/ (像素风头像.png, 像素公众号头图示例.png, 像素品牌资产.png) — these are the maintainer's (花叔 / @AlchainHust) personal IP assets, included as a working example of the "personal brand auto-inject" pattern.

For your own use, you have two options:

1. Replace: Swap in your own logo / avatar / style example PNGs and edit the trigger words in SKILL.md (search for "花叔风格")
2. Disable: Delete assets/personal-brand/ — the skill's main features (slides, single covers, brand asset protocol) all still work

What's bundled

huashu-slide-codex/
├── SKILL.md                   # Agent instructions (Chinese, but agent is bilingual)
├── test-prompts.json          # 7 test prompts covering all paths
├── assets/personal-brand/     # 3 pixel-style example assets (swappable)
├── references/
│   ├── design-principles.md
│   ├── prompt-templates.md
│   ├── proven-styles-gallery.md
│   ├── proven-styles-snoopy.md
│   └── design-movements.md
└── scripts/
    ├── create_slides.py       # PPTX assembly (python-pptx)
    ├── image_deck_html.py     # HTML deck assembly
    ├── html2pptx.js           # Path 3 HTML → PPTX conversion
    └── upload_image.py        # ImgBB upload (stdlib only)

All script references in SKILL.md use a [SKILL_DIR]/... placeholder. The skill is fully self-contained — no machine-specific absolute paths.

Iteration history

This skill went through 8 iteration rounds before this release, most driven by real-world delivery failures and successes. See results.tsv in darwin-skill for the full optimization log. Key learnings:

• R3 (full_test): When pixel-style was triggered but Codex put 花叔 character on a coffee mug instead of as the protagonist — added a "character must appear as scene protagonist" rule.
• R4 (full_test): For important single images, generating 3 differentiated versions × different design philosophies is always worth the extra 2 image_gen calls.
• R6 (full_test): AI defaults to filling every slide to the ceiling — added page-type-aware density tiers ("ceiling is not target").
• R7 (full_test): OpenClaw deck validated the Spread Frame + Mascot Continuity patterns.
• R8 (full_test): Codex rationalized switching to HTML mode by claiming "content needs precision" — added the 🔴 default-path-lock rule with explicit forbidden-justification patterns.

Related skills

• huashu-design — HTML-native design skill, works in any runtime
• huashu-skills — Other skills (AI proofreading, topic gen, image generation, etc.)
• darwin-skill — The autonomous optimization framework used to iterate on this skill

License

MIT — free for personal and commercial use, no authorization required.

Author

花叔 · @AlchainHust — AI Native Coder, indie developer, ~30 万粉自媒体. Made many AI skills, ships many AI products.