SKILL.md
Query: $ARGUMENTS
Goal: answer the query with minimum fetches. Targeted reads only — no broad crawling.
Sequence:
- PARSE — Extract from $ARGUMENTS:
- Target source type: github | url | mcp-tool | search - Specific artifact: endpoint, schema, function, config key, page section - If source is ambiguous → ask before fetching
- LOCATE — One targeted lookup to find the exact resource:
- GitHub repo: fetch https://api.github.com/repos/<owner>/<repo>/contents/<path> or raw file URL - API spec: fetch OpenAPI/AsyncAPI root, then only the relevant path object - DB schema: use available MCP tool (e.g. list-tables → describe only relevant tables) - Unknown source: one WebSearch with site: scoped query, pick top result
- FETCH — Read only what answers the query:
- For large docs: fetch the section anchor/fragment if URL supports it - For OpenAPI specs: extract only the relevant paths + components/schemas entries - For GitHub files: prefer raw URL over HTML — less noise, fewer tokens - Stop after 2 fetches if answer is found
- SYNTHESIZE — Output:
SOURCE: <url or tool:resource>
FINDING: <direct answer to the query — facts, not narrative>
SCHEMA: <relevant fields/types if applicable>
NOTES: <caveats, version constraints, gotchas>Hard rules:
- Never fetch a page just to find where the real page is — resolve first, then fetch
- No recursive crawling — if a link looks relevant, include it in NOTES, don't follow it
- Skip auth/login pages — note the access requirement and stop
- If MCP tool is available for the source (DB, internal API), prefer it over web fetch
- Max 3 fetches per query — if not resolved: report what was found + what's still missing