mcp-to-skill

name	mcp-to-skill
description	Use when a workflow built on an MCP server has stabilized and should become a production skill, or when the user says "convert this MCP workflow to a skill". Produces a skill with a bundled script that calls the underlying API directly, eliminating the per-session MCP token overhead.

MCP to Skill

MCP is the prototype surface; skills are the production surface. A connected MCP server preloads its tool definitions into every session - a heavyweight server costs 15-25K tokens before you type anything. A skill costs its frontmatter (~60 tokens) until invoked, and a bundled script costs near zero because deterministic work happens in code, not in context. The Lab 3 email labeler measured an order of magnitude fewer tokens after conversion.

Convert when: the workflow is repeated, its steps have stabilized, and the MCP server's tools map to a documented API. Don't convert one-offs or workflows still changing shape.

Procedure

1. Capture the working workflow. From the current session (or the user's description): which MCP tools get called, in what order, with what arguments, and what decisions happen between calls.
2. Map tools to API calls. Find the underlying service's API docs. Each MCP tool call becomes a direct HTTP request or official-client call. Note the auth model (API key, OAuth) and how the MCP server was authenticating; the script needs its own credentials path, usually an env var or a stored token file.
3. Split judgment from determinism. Everything mechanical (auth, pagination, retries, request shaping) goes in a script under the skill's scripts/ directory. Everything requiring judgment (what to do with the data, edge-case decisions) stays in SKILL.md prose. If the script needs the model's judgment mid-run, restructure: script fetches, model decides, script applies.
4. Write the script. Single file, stdlib or minimal deps, arguments via CLI flags, results as JSON to stdout, errors to stderr with nonzero exit. Credentials from env vars; document the names, never hardcode values.
5. Write SKILL.md per the write-a-skill conventions: frontmatter description with "use when" triggers, body that says when to run the script, how to interpret output, and what to do on common failures.
6. Test in a fresh session. Invoke the skill, run the real workflow end to end. Fix until it matches what the MCP version did.
7. Retire the server. Remove it from MCP config. Run /context before and after; report the savings as a number. That number is the point.