菜单
Turn any component site, asset library, or API into something agents can use — a registry entry, a skill, or a full MCP server. Load when the user says "add this site/library to Atelier" or wants to contribute a new integration.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Atelier's native asset tools — image→webp compression, video→webp frame sequences, scrollytelling scaffolding, and AI video generation (Gemini Omni Flash / Veo 3.1) with a graceful no-credentials fallback. Load when handling images, video, storytelling sections, or media generation.
The Atelier operating loop — understand → clarify → plan → execute → review → compare-vs-done-state. Governs when to stop, when to ask, which model tier runs each phase, and when to think deep or compact. Load for ANY nontrivial task.
The thinking protocol — load FIRST on every prompt, before any other skill. Makes any model (Opus, Sonnet, Haiku) reason like a principal engineer - decisive answers, root-cause debugging, security reflexes, outcome-first communication. This is step 0 of the loop.
The design-truth gateway — resolves which design files are authoritative right now (core law + self-learning overlays). Run at the START of any UI/design task, before reading the design law.
The Atelier Auth Law — production-hardened session architecture (rotating refresh tokens, families, grace window, silent refresh, sliding window) plus the anti-patterns that corrupt auth. Load before designing or touching ANY authentication code.
The Atelier Clean Code Law — 250-word-per-file budget, DRY via the graph, ponytail minimalism, and the comment policy. Load before writing or refactoring any source code.
| name | mcp-maker |
| description | Turn any component site, asset library, or API into something agents can use — a registry entry, a skill, or a full MCP server. Load when the user says "add this site/library to Atelier" or wants to contribute a new integration. |
Not everything needs an MCP server. Pick the cheapest rung that works:
| The source is… | Build this | Effort |
|---|---|---|
| Copy-paste components with a repo (ReactBits-style) | Registry row + the GitHub-MCP pattern (agent reads real source from the repo) | 2 min |
A library with a CLI (npx x add <c>) | Registry row with the install command | 2 min |
| Knowledge/patterns/rules (a design system, a guide) | Skill (SKILL.md) | 30 min |
| A live API the agent must query (search, generate, fetch) | MCP server | hours |
python <plugin>/db/store.py exec "INSERT INTO registry (kind,name,source,install,usage,meta)
VALUES ('component','<Name>','<url>','<install cmd>','<when/how to use>','<tags>')"
For repos without CLIs, set usage to: "read source via GitHub MCP: search / for " — the agent pulls real code, zero hallucination.
skills/<name>/SKILL.md: frontmatter description ≤2 sentences stating WHEN to
load it; body teaches decision rules and exact values, not tutorials. Add a
framework_map row so store.py map <topic> finds it. Follow clean-code-law
for any bundled scripts.
Minimal stdio server exposing 2–3 tools (search_x, get_x); TOON-lean
responses (no JSON bloat back to the model); version-pin in registry;
document env vars in .env.example. If an official MCP exists (shadcn,
aceternity), NEVER rebuild it — registry-row it.
Fork → add row/skill/server per above → self-check (skill loads, row queries, server answers) → PR using the proposal template. Review gates apply (ponytail: the ladder rung must be justified).