skill-composer

One install. Every AI agent on your machine now chains skills automatically.

Stop manually running /frontend-design then /animate then /polish then /audit. Skill-composer reads the prompt, picks the right chain, and runs the whole pipeline in order — across every AI agent it finds on your system.

Install (one shot)

npx skill-composer install

That's it. No flags, no config. It auto-detects every agent installed under your home directory and registers itself into each one — Claude Code, Codex, Cursor, Windsurf, Amp, Gemini, Cline, Continue, Aider, Roo, Qwen, Copilot, plus anything else you declare in composer.config.json.

To see which agents were auto-detected on your machine before installing:

npx skill-composer agents

To remove later:

npx skill-composer uninstall

What you get

• One install, every agent. Detects Claude Code, Codex, and Cursor in one pass.
• 13 prebuilt chains. Design Pipeline, Accessibility, Onboarding Flow, Plugin Development, Remotion Video, and more — already wired to the skills on your machine.
• Auto-detected chains. Scans skills you already have and infers new chains from category, trigger, and input/output overlap.
• Remote skill cache. Pulls SKILL.md files from skills.sh on demand, caches them for 14 days, and uses them same-turn without forcing a permanent install.
• No runtime overhead. The orchestration is one ~7 KB skill. Chains are precomputed at install time and embedded directly into the SKILL.md the agent loads.

How long does chaining take?

Chaining adds zero latency at prompt time:

• Install: under 1 second. One file written per agent (~30 KB each, depending on how many skills you already have).
• Per prompt: the agent reads the embedded chain map (already in its loaded skill context) — no extra tool calls, no extra round trips.
• Per skill in the chain: runs at the agent's normal speed. A 4-skill chain executes in roughly the same wall-clock time as you running each skill manually, except you didn't have to type four prompts and remember the right order.

Does it save tokens?

Yes — in two ways:

1. No skill-discovery prompts. Without skill-composer, you typically waste 200–500 tokens per turn asking the agent which skill to use, or pasting skill names. The chain map is preloaded, so the agent already knows.
2. No re-uploading SKILL.md. Cached remote skills are read locally on second use instead of re-fetched and re-tokenized from the network.

The orchestration skill itself costs roughly 2,000 tokens of context for the base routing rules, plus ~50–100 tokens per skill it indexes for you. On a typical install with 50 skills that's ~7,000–8,000 tokens, paid once per session — not per turn.

Real-world example

Before:

You: build a settings page
Agent: (uses generic frontend skill, output is bland)
You: now add animations
Agent: (re-reads context, adds basic animations)
You: now polish it
Agent: (...)
You: now audit accessibility
Agent: (...)

Four prompts. Four context reloads. Easy to forget the audit step.

After:

You: build a settings page
Agent: (chains frontend-design → animate → polish → audit automatically)

One prompt. The agent runs the full Design Pipeline because it recognizes the trigger.

Power-user commands

You don't need any of these for normal use, but they're there:

npx skill-composer agents            # list AI agents auto-detected on your machine
npx skill-composer scan              # list every skill found across your agents
npx skill-composer chains            # show the 13 prebuilt + auto-detected chains
npx skill-composer discover <query>  # rank skills relevant to a query
npx skill-composer remote fetch <skills.sh-url>   # cache a remote skill
npx skill-composer remote search <query>          # search cached remote skills
npx skill-composer graph             # visualize skill relationships

Supported agents

Auto-detected from a built-in registry. An agent is included only if its root directory exists on your machine — skill-composer will never write into agents you don't have installed.

| Agent | Detected at | |-------------|---------------------| | Claude Code | ~/.claude/skills | | Codex | ~/.codex/skills | | Cursor | ~/.cursor/skills | | Windsurf | ~/.windsurf/skills| | Amp | ~/.amp/skills | | Gemini | ~/.gemini/skills | | Cline | ~/.cline/skills | | Continue | ~/.continue/skills| | Aider | ~/.aider/skills | | Roo Code | ~/.roo/skills | | Qwen | ~/.qwen/skills | | Copilot | ~/.copilot/skills |

Need another agent? Add an entry under agents.<name> in your composer.config.json with at least a skillDirs array. It will be auto-detected and installed into just like the built-ins — no code changes required. The generic adapter handles any markdown skill format.

Run npx skill-composer agents to see what's currently detected on your machine.

Requirements

• Node.js 18+
• macOS or Linux (Windows path support is best-effort)

License

MIT — see LICENSE.