| name | Context Intelligence |
| description | Use when work may touch a codebase, long session, large files, unknown repo structure, or model context is limited. Finds relevant context first, avoids broad dumps, keeps a compact working map, and recommends compact/fresh-session points. |
| allowed-tools | Read, Bash |
Context Intelligence
Use this skill before large implementation, debugging, refactors, or unfamiliar repos.
Context Budget Rules
- Start with intent, not repo scan. Identify target feature, error, route, component, API, or user flow.
- If
AGENTS.md, CLAUDE.md, or docs/agent-memory/INDEX.md exists, read only the relevant boot/memory files first, then verify claims against code.
- Use
Bash with rg, rg --files, or find to find candidate files. Read only likely relevant files.
- Prefer file maps and symbol names over full file contents until edit scope is clear.
- Keep a short context ledger in the response: goal, relevant files, assumptions, decisions, next edit scope.
- If session history grows or repeated full tool context appears, recommend
/compact with a concrete focus.
- For unrelated tasks, recommend fresh session instead of carrying stale history.
Search Pattern
- Find entrypoints: routes, pages, controllers, handlers, package scripts.
- Find data model: schema, types, validators, API contracts.
- Find UI components: parent page, reusable components, styles, state.
- Find tests: nearest unit/e2e/integration coverage.
- Find config only when needed: build, lint, test, env, framework.
Output Standard
Return concise context map:
GoalLikely FilesRiskPlanNeed Read Next
Do not read huge logs or generated files unless user asks. Summarize long outputs.
