prompt-finder

A universal AI-agent skill / tool for finding curated prompt templates.

Defaults to a free public library at https://zhouyu.space (Mojo), but the backend URL is configurable per install. Any backend implementing the documented HTTP contract works — this is not specific to one project or one agent system.

• 📜 SKILL.md — Claude Code / Claude Desktop skill descriptor
• 📋 openapi.yaml — OpenAPI 3.1 spec for Codex / GPT function calling / Cursor / LangChain / any HTTP-aware agent
• 📖 API doc — see openapi.yaml for full contract

---

What it does

When the user asks for a starting-point prompt template (not "do the task for me"), the skill:

1. Probes the backend's category tree
2. Searches by category / subcategory / keyword / tool
3. Picks the best match
4. Returns the prompt content verbatim, along with title / summary / cover image (if any) / source attribution

Categories (default tree):

• 写作 / Writing — 公文 (formal docs) / 短篇 (short-form) / 随笔 (essay) / 长篇 (long-form)
• 图像 / Image — 电商 / 纯生成 / 图像编辑 / 优化
• 视频 / Video — 运镜 / 创意 / 电商
• 工程 / Code — Skill / PRD / 代码

---

Install per agent

Option A — Claude Code / Claude Desktop

Skill auto-discovery via ~/.claude/skills/:

mkdir -p ~/.claude/skills/prompt-finder
curl -sSL -o ~/.claude/skills/prompt-finder/SKILL.md \
  https://raw.githubusercontent.com/n8f86p7j5b-afk/prompt-finder/main/SKILL.md

Restart Claude. Skill appears in available skills automatically. Triggers: "find a prompt for X" / "给我一个 prompt" / "提示词模板".

Option B — OpenAI Codex / GPT Function Calling / GPT Actions

Import openapi.yaml as a custom GPT Action or as function-calling tool spec:

curl -sSL https://raw.githubusercontent.com/n8f86p7j5b-afk/prompt-finder/main/openapi.yaml \
  > prompt-finder.yaml
# Then in OpenAI Playground / Custom GPT builder → Actions → "Import from URL" or paste contents

GPT will auto-generate function signatures for the 5 endpoints (searchPrompts, getPrompt, randomPrompt, topPrompts, listCategories).

Option C — Cursor

Cursor supports MCP and .cursorrules. Quick way:

mkdir -p ~/.cursor/rules
curl -sSL -o ~/.cursor/rules/prompt-finder.md \
  https://raw.githubusercontent.com/n8f86p7j5b-afk/prompt-finder/main/SKILL.md

Or paste the SKILL.md content into your .cursorrules at the project root.

Option D — LangChain / LangGraph / autogen / any Python agent

import httpx

BASE = "https://api.zhouyu.space"  # configurable

def search_prompts(category=None, subcategory=None, q=None, page_size=10):
    params = {k: v for k, v in {
        "category": category, "subcategory": subcategory, "q": q,
        "page_size": page_size,
    }.items() if v is not None}
    r = httpx.get(f"{BASE}/api/skill/prompts", params=params, timeout=15)
    r.raise_for_status()
    return r.json()

# Register as a LangChain Tool, autogen function, etc.

For LangChain, wrap as @tool or import via OpenAPI loader. See the openapi.yaml.

Option E — Direct HTTP (curl)

No SDK needed. Test from any shell:

curl -s 'https://api.zhouyu.space/api/skill/categories'
curl -s 'https://api.zhouyu.space/api/skill/random?category=图像'
curl -s 'https://api.zhouyu.space/api/skill/prompts?category=图像&subcategory=电商&q=香水&page_size=5'

---

Configuring a different backend

Default base URL: https://api.zhouyu.space.

To use your own instance (e.g. a self-hosted Mojo or a compatible server):

• For Option A (SKILL.md): open ~/.claude/skills/prompt-finder/SKILL.md, search PROMPT_FINDER_BASE, replace.
• For Option B (openapi.yaml): add another entry under servers: and point your tool at that URL.
• For Option D (Python): just change the BASE constant.

Any server that implements the contract in openapi.yaml is plug-compatible.

---

API contract — <base>/api/skill/*

See openapi.yaml for the full spec. Five endpoints:

Method	Path	Use
GET	/categories	Probe the category tree
GET	/prompts?category=&subcategory=&q=&tool=&mode=&page_size=	Search
GET	/prompts/{id}	Fetch by id
GET	/random?category=	Random pick
GET	/top?limit=	Top N popular

Every prompt returns 17 fields (see Prompt schema in openapi.yaml):

id, title, title_zh, content, summary, summary_zh,
category, subcategory, tags, mode (creation|optimization),
language (zh|en), recommended_tool, tool_url, difficulty (小白|进阶|专业),
cover_image_url, source_url

Agents should prefer Chinese fields (title_zh / summary_zh) when the user speaks Chinese.

---

Why use a curated library (vs the agent generating prompts on the fly)

• Curation > generation: The default backend's editorial rubric rejects "帮我写 X" one-shot requests; what you get back is a real template with role / structure / output constraints.
• Reproducibility: Templates have placeholders (【主题】, {role}) you adapt — same template gives same shape of output across runs.
• Attribution: Each result links back to the source page with author info.
• Reference images: Image/video prompts include cover_image_url so you see what the prompt produces.

---

Default backend — Mojo / zhouyu.space

• Public, free, anonymous. No API key.
• Rate limit: 60 req/min per IP.
• ~1300 curated entries (target), categorized and LLM-audited.
• Edited per a strict "prompt template, not request" rubric.
• Free to use in any project, commercial or otherwise. Please credit Mojo (zhouyu.space) if you publish a derivative.

---

License

Skill descriptor / OpenAPI spec: MIT.

Library content licensing varies per entry — each prompt's source_url / platform_author field links to original attribution. Default backend (Mojo) is free public access.