claude-enforce

Transform CLAUDE.md rules into deterministic Claude Code hooks. Automatically.

  ClaudeEnforce — Init

  Parsing CLAUDE.md...
  Found 8 rules in 3 sections

  Classification Results:

    Enforceable (hook-ready):  5
    Advisory (CLAUDE.md only):  3

    ✓ [CRITICAL] NEVER edit .env files
    ✓ [CRITICAL] NEVER commit secrets or API keys
    ✓ [HIGH]     ALWAYS run tests before committing
    ✓ [HIGH]     NEVER use git push --force
    ✓ [HIGH]     Do not skip git hooks with --no-verify
    · [NORMAL]   Keep functions under 50 lines
    · [NORMAL]   Use TypeScript strict mode
    · [NORMAL]   Prefer composition over inheritance

  Generated 5 hooks → .claude/settings.json

CLAUDE.md is advisory. Hooks are deterministic. Claude can ignore your CLAUDE.md under pressure. It cannot bypass a hook that exits with code 1. This tool bridges the gap.

Try it

npx claude-enforce init              # parse CLAUDE.md → generate hooks
npx claude-enforce install --list    # browse pre-built presets
npx claude-enforce audit             # compliance report from transcripts

Commands

init — Parse rules, generate hooks

claude-enforce init                  # reads ./CLAUDE.md
claude-enforce init --file path.md   # custom CLAUDE.md path
claude-enforce init --dry-run        # preview without writing

Reads your CLAUDE.md, classifies each rule as enforceable (can be a hook) or advisory (stays in CLAUDE.md), and generates .claude/settings.json hooks for enforceable rules.

Set ANTHROPIC_API_KEY for AI-powered classification. Without it, pattern-based classification works fine.

install — Pre-built presets

claude-enforce install --list
claude-enforce install block-destructive-bash no-env-edit

| Preset | Hooks | What it blocks | |--------|-------|---------------| | block-destructive-bash | 3 | rm -rf, git reset --hard, force push, DROP TABLE | | no-env-edit | 2 | Any Edit/Write to .env files | | no-commit-without-tests | 1 | git commit unless tests were run | | no-migration-without-confirm | 1 | Database migrations without approval | | typecheck-on-save | 2 | Runs tsc --noEmit after editing .ts files |

audit — Compliance report

claude-enforce audit                 # scan .claude/logs/
claude-enforce audit --json          # machine-readable
claude-enforce audit --since 2025-01-01

  Compliance Report
  ──────────────────────────────────────────

  Tool calls analyzed:   247
  Total violations:      12
  Compliance rate:       95.1%

  Most Broken Rules:
    4x no-destructive-bash
    3x no-force-push
    2x no-env-edit

watch — Real-time monitoring

claude-enforce watch                 # live monitor
claude-enforce watch --interval 5000

Polls .claude/logs/ for new transcript entries and alerts instantly on violations.

How hooks work

CLAUDE.md = instructions Claude reads. Hooks = shell commands that intercept tool calls.

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "type": "command",
        "command": "if echo \"$CLAUDE_TOOL_INPUT\" | grep -qE \"rm -rf\"; then echo \"BLOCKED\" >&2; exit 1; fi"
      }
    ]
  }
}

• PreToolUse: runs before a tool. Exit 1 = tool blocked.
• PostToolUse: runs after. Used for checks (lint, typecheck).
• $CLAUDE_TOOL_INPUT: JSON of what Claude is trying to do.

Part of the Claude Code Toolkit

| Tool | What it does | |------|-------------| | claude-score | Score your setup A-F, see what's missing | | claude-gen | Generate CLAUDE.md from your codebase | | claude-enforce | Convert rules into deterministic hooks |

npx claude-score         # diagnose
npx claude-gen           # fix
npx claude-enforce init  # protect

License

MIT