claude-foundations/improvements/hooks-and-automation.md

Hooks & Automation

Why This Matters

Hooks are the #1 "why didn't I do this sooner" recommendation across the community. Unlike CLAUDE.md rules (which Claude can forget mid-session), hooks always execute. They provide deterministic guardrails around non-deterministic AI behaviour.

Key principle: "Never send an LLM to do a linter's job. LLMs are expensive and slow compared to traditional linters."

Current State

• No hooks configured in ~/.claude/settings.json
• Formatting/linting rules live in CLAUDE.md as prose instructions

Community Recommendations

PostToolUse: Auto-Formatting

Run formatters automatically after every Edit/Write tool call. Eliminates all formatting rules from CLAUDE.md.

• prettier --write for JS/TS/JSON/YAML/MD
• shfmt -w for shell scripts
• ruff format for Python
• Plugin: ryanlewis/claude-format-hook supports multi-language detection

Configuration goes in ~/.claude/settings.json under hooks.PostToolUse.

PreToolUse: Security Gates

Block dangerous operations before they execute:

• Block writes to .env, .key, .pem, secrets/, lockfiles
• Block destructive commands (rm -rf /, dd, mkfs)
• Scan for API keys/credentials in file content before writes
• Block commits that include sensitive file patterns

PreToolUse: Commit Validation

Wrap Bash(git commit) with test/lint validation. Only allow commits if checks pass. Forces Claude into "test-and-fix" loops until the build is green. This is considered the single most effective quality gate.

Important: Don't block at write time — blocking on Edit/Write confuses Claude mid-plan. Validate at commit time instead.

PreCompact: Transcript Backup

Save conversation transcript before auto-compaction so context is never lost. Creates a timestamped backup in a known location. Pairs well with the auto-memory system.

Reference: https://yuanchang.org/en/posts/claude-code-auto-memory-and-hooks/

Notification Hook

Desktop notifications when Claude finishes a long task or needs input. Useful when running background agents.

Plugin: CC Notify from awesome-claude-code.

Anti-Patterns

• Don't add too many hooks — each one adds latency to every tool call
• Don't block on Edit/Write for linting — do it on PostToolUse (non-blocking) or at commit time
• Don't duplicate hook logic in CLAUDE.md — if a hook handles it, remove the prose rule