Skills
Skills are Markdown files rooted at .claude/skills/*/SKILL.md. Agents load them selectively through Anthropic-style Skill tool calls rather than injecting every guideline into each phase—keeping token curves predictable.
Categories (conceptual, not exhaustive)
| Category | Examples |
|---|---|
| Planning | Feature decomposition, sizing matrices, naming guidance |
| Testing | Unit / integration / contract / e2e methodology indexes |
| Language conventions | Go, TS, JVM, Rust, Ruby, Python, .NET idioms tenants expect |
| Review lenses | Cross-cutting checklist, dependency hygiene, observability deltas |
| Campaign | Child contract authoring, rollout coordination patterns |
Skills can depend on anchors such as #contract-tests headings so reviewers and coders converge on vocabulary.
Authoring overlays
Skills respect layer merge semantics for non-memory paths—that is file replacement:
- Extend base by adding sibling directories under
.claude/skills/<name>/in your tenant overlay or repo.coro/. - To override packaged guidance wholesale, publish a SKILL with the identical directory name; last-wins semantics ensure your version is the one materialised.
Keep skills actionable (checklists, anti-patterns, links to internal ADRs) rather than narrative marketing text; long essays belong in memory/ or external docs.
Practical guidance for proposal mechanics lives in Self-improvement when you need to open a PR for skill updates.