Update Claude Docs

CLAUDE.md Updater

Synchronize CLAUDE.md files based on staged git changes, capturing conventions and gotchas that Claude cannot infer by reading code.

When to Activate

• After significant code changes that reveal non-obvious patterns or constraints
• User requests a documentation sync or CLAUDE.md update
• Staged changes include new commands, env vars, or architectural decisions
• A new gotcha or non-obvious behavior has been introduced

Content Model

CLAUDE.md files should contain conventions, gotchas, and team preferences — not structural inventory. The guiding question for each potential entry is: "Would removing this cause Claude to make a mistake?" If not, omit it.

Include:

• Non-obvious bash commands (env vars required, non-standard behavior, ordering constraints)
• Code style rules that differ from language defaults
• Testing instructions and non-standard runner configuration
• Repository etiquette (branch naming, PR conventions)
• Architectural decisions specific to this project
• Developer environment quirks (required env vars, local service dependencies)
• Common gotchas — behaviors that regularly surprise contributors

Exclude:

• File-by-file directory listings (Claude can use git ls-files)
• Dependency version tables (Claude can read package.json)
• Standard language or framework conventions Claude already knows
• [TODO] placeholder entries
• Structural overviews Claude can derive from code

Quick Process

1. Check Git: Verify in git repository; determine interactive vs. automated mode
2. Get Staged Files: git diff --cached --name-status
3. Group by CLAUDE.md: Find nearest CLAUDE.md ancestor for each staged file
4. Analyze Changes: Ask "what does this change reveal that is non-obvious?"
5. Deduplication Check: Verify proposed content is not already in ancestor CLAUDE.md files
6. Propose .claude/rules/ Factoring: Suggest a rules file if topic warrants it
7. Show Summary: Preview proposed additions (interactive mode only)
8. Apply Updates: Write changes after confirmation (or conservatively in automated mode)

Update Triggers

An update is warranted when:

• A new non-obvious convention or pattern is introduced
• A new command, env var, or setup step is added that is not self-evident
• An existing documented behavior changes in a surprising way
• A new package with cross-cutting concerns is added
• Significant architectural changes (>50 lines) that introduce non-obvious constraints

Skip updates for: formatting changes, typos, renames with no behavior change, or anything that is standard language/framework convention.

Output Targets

• CLAUDE.md files: Primary output — update nearest ancestor CLAUDE.md for each change
• .claude/rules/<topic>.md files: Secondary output — when a topic-specific rule is identified that would push CLAUDE.md over 200 lines or only applies to a file subset, propose factoring it into a rules file with optional paths frontmatter

Length Constraint

Target: under 200 lines per CLAUDE.md. If an update would push a file over 200 lines, propose factoring detailed content into a .claude/rules/<topic>.md file instead.

Automated Mode (Non-Interactive)

When running without a human present (e.g., via a pre-commit hook):

• Only append content that is clearly missing and high-value
• Prefer a no-op over a low-quality update
• Do not ask questions or prompt for confirmation
• Log what was added (or that no update was made)

Safety

• Git provides rollback (git restore CLAUDE.md)
• Single confirmation prompt in interactive mode
• Review with git diff **/*CLAUDE.md before committing
• Outdated entries may be pruned when a significant update is made

Best Practices

1. Stage changes first (git add)
2. Review updates (git diff **/*CLAUDE.md)
3. Commit CLAUDE.md changes together with related code
4. Treat CLAUDE.md as a living document — prune outdated entries over time