doctor

name	doctor
description	Validate your .claude/ configuration. Checks that commands work, paths exist, hooks are valid, and agents are well-formed. Use after /claude-init or /update to catch issues.
allowed-tools	Read, Bash, Glob, Grep

Doctor: Validate Your Config

Run diagnostics on the .claude/ configuration and report issues.

Checks

1. CLAUDE.md Health

• Exists at repo root
• Under 80 lines (warn if over, error if over 200)
• All @import paths resolve to existing files
• Commands section: verify each command actually runs (e.g., npm test --help, cargo test -- --help)
• No placeholder text left from templates (e.g., <your-framework-here>)
• No content that merely restates what Claude could derive by reading the codebase (full dependency/version lists duplicating package.json, directory trees ls would show, verbatim README paragraphs) — flag for trimming, matching Claude Code's own /doctor CLAUDE.md-trimming check (changelog 2.1.206)

2. Agents

For each .claude/agents/*.md:

• Valid YAML frontmatter (name, description present)
• Model is valid (opus, sonnet, haiku, fable, inherit, or a full model ID)
• Referenced tools exist
• background is a boolean and memory is one of user/project/local if present
• No empty markdown body

3. Skills

For each .claude/skills/*/SKILL.md:

• Valid YAML frontmatter
• name matches the skill's directory name and is lowercase-kebab-case
• Description is specific enough (not just "does stuff")
• Description is third-person and contains a when-to-use trigger (Use when…, Use before…, or Auto-triggered…) — descriptions with no trigger won't reliably fire
• Description is ≤1024 characters
• Description does NOT summarize the workflow steps — a process summary in the description makes Claude follow the summary instead of reading the skill
• Skill body has Red Flags / rationalization-rebuttal content and a verification step (the generated skills should all carry these)
• SKILL.md body is under 500 lines (warn over ~150 — push long examples/scripts into sibling reference files)
• If context: fork, verify no Task/Skill tool references in body (forked skills can't spawn subagents)
• If paths: frontmatter exists, verify globs match at least one file (dead scopes = skill never loads)
• allowed-tools uses real tool names (Read, Edit, Write, Bash, Grep, Glob, Agent — not Tool, File, etc.)
• Any !`cmd` or ```! shell-preprocessing blocks run read-only commands only (never rm, curl http://..., writes, or long-running ops)
• Referenced scripts in scripts/ directory exist and are executable

4. Rules

For each .claude/rules/*.md:

• If paths: frontmatter exists, verify those glob patterns match at least one file
• No duplicate rules across files
• Content is actionable (not just "write good code")
• No dangling symlinks under .claude/rules/ — find .claude/rules -xtype l should return nothing (shared rule libraries are linked in via symlink; a broken link silently drops those rules from context)

5. Hooks

Read .claude/settings.json:

• Valid JSON (strict — "//" comment keys are OK, but trailing commas and JS comments break Claude Code's parser)
• Hook event names are real — current set (per https://code.claude.com/docs/en/hooks): SessionStart, SessionEnd, Setup, UserPromptSubmit, UserPromptExpansion, PreToolUse, PermissionRequest, PermissionDenied, PostToolUse, PostToolUseFailure, PostToolBatch, Notification, MessageDisplay, SubagentStart, SubagentStop, TaskCreated, TaskCompleted, Stop, StopFailure, TeammateIdle, InstructionsLoaded, ConfigChange, CwdChanged, FileChanged, WorktreeCreate, WorktreeRemove, PreCompact, PostCompact, Elicitation, ElicitationResult — a typo like preToolUse or OnEdit silently never fires
• Hook handler type is present and one of command, http, mcp_tool, prompt, agent (required field — no default)
• Hook commands are executable (if type: command)
• No overly broad matchers that would slow every tool call
• Deny patterns don't block normal development workflows
• PreToolUse hooks use the current decision schema (hookSpecificOutput.permissionDecision), NOT the legacy {"decision": "allow"|"block"} form, which PreToolUse ignores
• Regex matchers are anchored with word boundaries where needed (e.g., git\s+push\s+(-f\b|--force\b) not git push.*-f — the lazy form matches git push-to-origin --forward)
• MCP matchers (if present) use regex form with .* (e.g., mcp__memory__.*, not mcp__memory)
• PostToolUse formatters use || true or exit 0 so a missing tool doesn't block writes
• Hook self-protection exists: .claude/settings.json, .claude/hooks/**, .claude/agents/**, .claude/skills/**/SKILL.md should be blocked from Edit/Write

5b. Output Styles (if present)

For each .claude/output-styles/*.md:

• Valid YAML frontmatter with name and description
• keep-coding-instructions is a boolean (not a string) if present
• The file is referenced by outputStyle in .claude/settings.json OR documented in CLAUDE.md so users know it exists

6. ARCHITECTURE.md

• Exists at repo root (warn if missing — /claude-init should have generated it)
• Under 120 lines (warn if over)
• No placeholder text from templates (e.g., <What this system does>, <ASCII diagram>)
• Directory Structure section matches actual top-level directories (compare against ls)
• Entry Points section references files that actually exist
• CLAUDE.md references it (should contain "ARCHITECTURE.md" somewhere)

7. Git Integration

• .claude/settings.local.json is in .gitignore
• .claude/agent-memory-local/ is in .gitignore
• .claude/checkpoints/ is in .gitignore (PreCompact hook writes here)
• .claude/settings.local.json.example IS committed (it's the onboarding artifact)
• No secrets in any .claude/ files (grep for API keys, tokens, passwords)

Output

## Doctor Report

Passed: N checks
Warnings: N
Errors: N

### Errors (must fix)
- [file] Description

### Warnings (should fix)
- [file] Description

### All Clear
- [file] OK

Rules

• Run every check, don't skip on first failure
• Be specific about how to fix each issue
• Don't modify any files — only report