🤖 ai kodi
9 Skill Categories for Claude Code (from Anthropic's Internal Use)
From Thariq’s thread on how Anthropic internally categorizes Claude Code skills into 9 types — plus practical tips for writing effective ones.
The 9 Skill Categories
| # | Category | Examples |
|---|---|---|
| 1 | Library/Tool Reference | billing-lib, internal-cli, frontend-design |
| 2 | Verification | signup-flow-driver, checkout-verifier |
| 3 | Data & Monitoring | funnel-query, grafana, cohort-compare |
| 4 | Workflow Automation | standup-post, create-ticket, weekly-recap |
| 5 | Scaffolding | new-workflow, new-migration, create-app |
| 6 | Code Quality | adversarial-review, code-style, testing-practices |
| 7 | Deployment | babysit-pr, deploy-service, cherry-pick-prod |
| 8 | Debugging/Incident | oncall-runner, log-correlator, service-debugging |
| 9 | Maintenance/Ops | resource-orphans, dependency-management, cost-investigation |
Key Tips for Writing Skills
- Skills are folders, not just markdown — Include scripts, assets, reference code, templates
- Gotchas section = highest signal — Built from real Claude failure points
- Progressive disclosure — Split into
references/,assets/, etc. Tell Claude what’s there - Be flexible, not rigid — Give info but let Claude adapt
- Description = trigger condition — Not a summary, but “when to use this”
- Config.json for user setup — Store per-user settings like Slack channels
- Skills can have memory — Log files, JSON, SQLite. Use
${CLAUDE_PLUGIN_DATA}for stable storage - Give Claude scripts — Code > prose for repetitive tasks
- Skill-scoped hooks — e.g.
/carefulblocks dangerous commands only when activated
The takeaway: Skills aren’t just documentation — they’re mini-applications. The best ones combine markdown guidance with executable scripts, persistent memory, and smart trigger conditions. Think of them as giving Claude a specialized toolkit, not just instructions.