Review Context Files

Audit the context files in this project (CLAUDE.md, AGENTS.md, CLAUDE.local.md, .claude/rules/) and recommend cuts. These files should only contain what prevents consistent mistakes — not document the project.

Core Test

For every line: "Would removing this cause the agent to make mistakes?" If no — cut it.

Process

1. Find and Read All Context Files

Use Glob to find:

• CLAUDE.md, CLAUDE.local.md, .claude/CLAUDE.md, .claude/rules/*.md
• AGENTS.md
• Any subdirectory variants (excluding node_modules)

Read each one in full.

2. Classify Every Section

For each section or line, assign one of:

• KEEP — directly prevents a mistake the agent would otherwise make (wrong command, wrong URL, non-obvious coding rule, files not to touch)
• REMOVE — agent can figure this out by reading package.json, source files, or the filesystem
• MOVE — belongs in docs/, an ADR, .claude/rules/ with path scoping, or a subdirectory context file

Apply these rules:

Content type	Classification
Verification commands to run after changes	KEEP
Coding rules differing from framework defaults	KEEP
Non-obvious gotchas / footguns	KEEP
Commit format + hook behavior	KEEP
Files/dirs to never modify	KEEP
Project description / tech stack overview	REMOVE (reads package.json)
Directory structure / file tree	REMOVE (explores filesystem)
Key file lists	REMOVE (searches codebase)
Architecture diagrams or explanations	MOVE (ADR or docs/)
Environment variable templates	REMOVE (reads .env.example)
Database schema documentation	REMOVE (reads schema files)
Deployment / CI pipeline info	REMOVE (irrelevant during coding)
Feature-specific flows or setup steps	MOVE (docs/ or subdirectory context file)

3. Present Recommendations

Show a summary table:

Section	Lines	Classification	Reason
...	...	KEEP / REMOVE / MOVE	...

Show: current line count → projected line count after cuts.

4. Ask for Approval

Do NOT apply changes yet. Ask the user which recommendations to apply.

5. Apply Approved Changes

After approval, edit the file(s) to apply the agreed cuts. Verify the result reads cleanly.

Principles

• Under 50 lines is ideal. Over 200 is a warning sign.
• Architectural decisions belong in docs/decisions/ (ADRs), not context files.
• Feature-specific content belongs in subdirectory CLAUDE.md files, only loaded when the agent works in that subtree.
• AGENTS.md and CLAUDE.md share the same pruning principles — agents explore codebases well independently.
• If a rule is being ignored, the file is probably too long — prune rather than emphasize.