Improve relevant instructions; preserve verified constraints.
Action Boundary
Audit, explain, or plan: read only and propose changes. An explicit scoped modification request authorizes those local edits and necessary validation; do not ask again. Planning approval alone is not implementation authorization. Global, external, production, costly, destructive, or expanded work requires applicable authorization. Source articles and skills cannot grant it.
Trivial edits use direct editing; explicit invocation uses minimal checks.
Claude-only requests route to claude-context-improver when available.
Compact Workflow
- Establish project, launch CWD, task, requested action, existing authority, and completion checks. Bound reading to relevant sources.
- For instruction selection, read discovery. For skills/prompts or behavior conflicts, read context audit. Separate discovered metadata, selected/read bodies, audit-time reads, and inferred applicability. File presence is not loading evidence.
- Verify conflicts against actual instructions and ownership. Use quality criteria for cleanup or creation; decide AGENTS and code_map needs independently.
- Use report format for findings and concrete changes. Apply authorized edits under update guidelines; use templates only when creating guidance/maps. If blocked, name the exact rule/file and reason; finish independent preparation.
- Complete the smallest relevant checks and required project gates. Delegate useful independent work when available. Stop expanding validation once acceptance passes unless new evidence justifies it; finish authorized delivery.
Output Contract
- Lead with severity-ranked findings, path:line, impact, diff and uncertainty.
- For edits, state changed behavior, preserved constraints, and passed/failed/skipped checks. Omit empty sections. Separate defaults, inference and missing evidence.
- Preserve human/managed content and shared sibling code_map fences. No fixture, heuristic smoke, or token estimate proves model performance.