Skills

The catalog. Each skill has its own page with a summary, its key opinions, install snippets, and a link to the canonical SKILL.md.

Currently shipping

  • rest-api-design: design and review HTTP REST APIs. Resource-oriented URLs, PATCH for state transitions, domain-expressive error codes, flat error envelopes, idempotency, content-type negotiation, typed contracts across TS/Python/Go/Rust.
  • structured-code-review: a review-only output format for code reviews. Source-of-truth-aware preamble, severity-tagged findings, file:line citations, no-findings-still-formal. Composes with domain-review skills.
  • task-handoff-summaries: three structured report formats. Implementation summary (before commit), worker handoff summary (multi-agent to orchestrator), closeout summary (after completion). Hard rules against using the summary to mask incomplete work.
  • cross-agent-review: workflow for routine cross-vendor agent peer review (Claude reviews Codex’s work; Codex reviews Claude’s). The handoff package with self-assessment redacted, the cold-review discipline, the disagreement protocol, the bounded iteration loop.
  • multi-agent-git-workflow: git discipline for multi-agent work. Branch hierarchy, worktree-per-agent topology, orchestrator/worker roles, merge authority, acceptance/rejection rules, plus universal commit discipline (Conventional Commits, mandatory task ID, co-author line, UAT gate, no silent amends).
  • branch-promotion-discipline: the layer above multi-agent-git-workflow. 3-tier develop to uat to main promotion, UAT branch as a long-lived environment, per-tier CI gate matrix, source-ref enforcement, hotfix flow with forward-merge, branch protection ruleset, pre-commit hook setup.

How skills are evaluated

Every skill in this collection passes through an evaluation loop before it ships:

  1. Draft the SKILL.md based on real policies or design intent
  2. Write 4 test prompts in evals/evals.json that probe the skill’s distinct opinions
  3. Spawn parallel runs, one with the skill loaded, one baseline without
  4. Grade outputs against per-case assertions; aggregate into a benchmark
  5. Iterate until the with-skill output is materially better than the baseline on the assertions that matter

The eval set ships in the repo at plugins/<plugin-name>/skills/<skill-name>/evals/evals.json so anyone can re-run it.

Contributing a new skill

To add a new skill to the collection:

  1. Plugin manifest: plugins/<plugin-name>/.claude-plugin/plugin.json with name, description, version.
  2. Skill file: plugins/<plugin-name>/skills/<skill-name>/SKILL.md in the standard SKILL format (YAML frontmatter with name, description; markdown body).
  3. Marketplace entry: add to .claude-plugin/marketplace.json under plugins with source: ./plugins/<plugin-name>.
  4. Evals (optional but the whole point): plugins/<plugin-name>/skills/<skill-name>/evals/evals.json with test cases and per-assertion pass/fail criteria.
  5. Docs page: skills/<skill-name>.md (this directory) with the human-facing summary, key opinions, install snippets, and a link to the SKILL.md.

Open a PR. If you ran evals, drop the iteration-1 benchmark in the PR description.

See the existing rest-api-design and structured-code-review pages for the format.

Table of contents