A skill for turning a high-performing reference image into an image-generation prompt for your own new content — by extracting why the reference works and carrying that hook onto your topic, instead of blindly copying its look.
Built for marketing / content workflows. Made by 发发.
把下面这句发给你的 coding agent(如 Claude Code),它会自动从 GitHub 拉取并安装:
帮我安装这个 skill:
git clone https://github.com/itsraefafa-crypto/fafa-hook ~/.claude/skills/fafa-hook,装完我重启会话。
或者直接在终端运行:
git clone https://github.com/itsraefafa-crypto/fafa-hook ~/.claude/skills/fafa-hook然后重启 Claude Code 会话即可使用。它遵循 SKILL.md 开放标准,在任何支持该标准的 agent(Claude Code、Codex、Cursor 等)中均可使用。
You give it a viral reference image (an infographic, a puzzle, a ranked list, etc.) and your new topic. It:
- Figures out why the reference grabs attention — its core mechanism, separated from its surface look.
- Negotiates a few choices with you (how closely to follow the look, who provides the copy, on-image text language, platform/size, target tool, number of variants).
- Produces a clean image-generation prompt you paste into your own image model.
It produces a prompt, not a picture. You take that prompt and run it in your own image tool (Doubao / Jimeng / Midjourney / etc.). Parts that must be exact — letter grids, precise data, your brand logo — are handed back as a separate compositing layer that you overlay in a layout/design tool, because image models cannot render those reliably. Think of it as a prompt-and-plan generator, not a one-click image maker.
Every viral image is core (the reason it works) plus skin (how it looks). The one test that separates them:
Remove the element. Is the image still compelling? If it collapses, it's core (always keep). If it still works, it's skin (keep, adapt, or rebuild — your choice).
The skill always preserves the core, and lets you decide how much of the skin to carry over.
- Drop the
fafa-hook/folder into your agent's skills directory. - Trigger it with something like "用这张参考图,帮我做一张关于〔你的主题〕的图" and attach the reference image.
- Answer a few short grouped questions.
- Receive the prompt + (if needed) a compositing note. Paste the prompt into your image tool; overlay any precise layer yourself.
| Decision | Meaning |
|---|---|
| content source | Do you provide the copy, or should it be devised? (If you provide it, it's used verbatim.) |
| visual | keep = same look, new content · adapt = same lineage, adjusted mood · reinvent = hook only, brand-new look (default keep) |
| platform / aspect | Output ratio, set by where you'll publish |
| text language | Language of the text rendered in the image — kept to one language, never mixed |
| image tool | Which model you'll paste into, so phrasing fits |
| variants | How many versions you want |
| logo (optional) | Upload a transparent-background logo + position; it's composited, never AI-generated |
- On-image text is one language only. The prompt is written in Chinese (for the model), but the text shown in the image follows your chosen language exactly, with no mixing.
- The prompt speaks only to the model. No explanations or annotations leak into the prompt — anything you need to know is told to you separately, in the conversation.
fafa-hook/
├── SKILL.md the skill itself
└── references/
├── interactive.md behavior-driven hooks (word-search, spot-the-difference, quizzes)
├── informational.md information-driven hooks (itineraries, checklists, rankings)
└── other-mechanics.md emotion / identity / suspense (starter placeholders to expand)
The reference files are meant to grow. As you analyze more viral formats in your domain, add their decomposition into the matching file — especially the organizing-logic library in informational.md and the placeholder mechanics in other-mechanics.md. That domain knowledge is what makes this skill yours.
MIT