go-skill-extractor

go-skill-extractor

A meta-agent that finds repeated Go patterns in a codebase and proposes them as new gophers skills. It does not create the SKILL.md on disk — it returns a draft proposal markdown the maintainer can iterate on.

When to Dispatch

Dispatch when:

• A service has grown past 50 packages and the team notices "we always do X"
• A monthly / quarterly audit on a long-lived monorepo
• After a large refactor that introduced new internal conventions
• A reviewer keeps copy-pasting the same nit comment across PRs

Do not dispatch when:

• The codebase is < 10 packages (not enough signal)
• The repeated pattern is just a function helper (factor it into a Go helper, not a skill)
• The pattern is project-specific (e.g., one company's auth flow) — those are runbooks, not skills
• An existing gophers skill already covers it (check the Skills Catalog first)

Input Contract

Input	Format	Example
module_root	path	./
min_occurrences (optional)	integer	3 (default)
min_packages (optional)	integer	2 (default)
existing_skills_dir (optional)	path	../gophers/skills (defaults to comparing against the Skills Catalog)
domain (optional)	comma-separated topic hints	errors,logging,http

Process

1. Enumerate packages.

   go list -f '{{.ImportPath}} {{.Dir}}' ./...
   

2. Detect candidate patterns. Look for structurally repeated shapes (≥ min_occurrences across ≥ min_packages packages):
   • Error patterns — repeated wrapping shapes, sentinel definitions, custom error types
   • HTTP patterns — middleware ordering, response shapes, request-id propagation
   • Logging patterns — repeated logger field sets, log-and-return mistakes
   • Concurrency patterns — same worker-pool shape, same select-with-context skeleton
   • Init patterns — repeated config loading, repeated DI wiring
   • Testing patterns — repeated fake shapes, repeated httptest harnesses
3. For each candidate, run negative coverage check. For every existing skill in existing_skills_dir:
   • grep -i <candidate-keyword> skills/*/SKILL.md
   • If any hit, skip the candidate (already covered)
4. Score candidates by (occurrences × packages × novelty). Return the top 3.
5. For each surviving candidate, draft a SKILL.md proposal following the gophers template — see Output Contract below.
6. Cite real file:line locations for every claim (this is not a hypothetical proposal — it must reference actual code).