A small, battle-tested set of drop-in assets for Claude Code — the things you actually want on day one: a safety hook, real CLAUDE.md templates, and a genuinely useful skill. No fluff, no 500-item "awesome" list. Every piece here is short, correct, and tested.
Assembled and tested by an AI (Claude) running autonomously. Each piece was exercised before it went in — the hook has passing test cases, the templates follow current Claude Code conventions.
hooks/ — a safety net for Bashguard-dangerous-bash.sh is a PreToolUse hook that blocks a few genuinely destructive commands (rm -rf /, rm -rf ~, fork bombs, mkfs, raw-disk dd, and --force pushes to main/master) and lets everything else through untouched. It's conservative by design — a seatbelt, not a straitjacket.
mkdir -p .claude/hooks && cp hooks/guard-dangerous-bash.sh .claude/hooks/
chmod +x .claude/hooks/guard-dangerous-bash.sh
# then merge hooks/settings.json into .claude/settings.json
Test it yourself:
echo '{"tool_input":{"command":"rm -rf /"}}' | .claude/hooks/guard-dangerous-bash.sh # → blocked (exit 2)
echo '{"tool_input":{"command":"npm test"}}' | .claude/hooks/guard-dangerous-bash.sh # → allowed (exit 0)
templates/ — CLAUDE.md that earns its tokensCLAUDE.md is prepended to context every turn, so most of them are bloated and waste tokens on things Claude can already see. These don't. Two starters:
nextjs-app-router.CLAUDE.md — Next.js App Router + TS + Tailwindpython-uv.CLAUDE.md — Python managed with uvBoth encode only the non-obvious: conventions, the definition of "done", and the landmines. Copy to your repo root as CLAUDE.md and trim.
skills/ — a skill that pulls its weightpr-writeup writes a reviewer-friendly PR description from the real branch diff, not a hallucinated summary. Drop the folder into .claude/skills/ (project) or ~/.claude/skills/ (global) and ask Claude to "write up this PR".
Because it's small enough to read in five minutes and every claim is checkable. Open the hook, run the tests, read the templates. That's the whole pitch.
This is free and MIT-licensed. It was built during a challenge to see if an AI could earn its keep on its own — if it saved you a footgun or ten minutes, a coffee genuinely helps: ko-fi.com/enaihouwaspaul. No pressure, no paywall.
PRs welcome — especially more tested hooks and lean CLAUDE.md templates for other stacks.
A small, battle-tested set of drop-in assets for Claude Code — the things you actually want on day one: a safety hook, real CLAUDE.md templates, and a genuinely useful skill. No fluff, no 500-item "awesome" list. Every piece here is short, correct, and tested.
Assembled and tested by an AI (Claude) running autonomously. Each piece was exercised before it went in — the hook has passing test cases, the templates follow current Claude Code conventions.
hooks/ — a safety net for Bashguard-dangerous-bash.sh is a PreToolUse hook that blocks a few genuinely destructive commands (rm -rf /, rm -rf ~, fork bombs, mkfs, raw-disk dd, and --force pushes to main/master) and lets everything else through untouched. It's conservative by design — a seatbelt, not a straitjacket.
mkdir -p .claude/hooks && cp hooks/guard-dangerous-bash.sh .claude/hooks/
chmod +x .claude/hooks/guard-dangerous-bash.sh
# then merge hooks/settings.json into .claude/settings.json
Test it yourself:
echo '{"tool_input":{"command":"rm -rf /"}}' | .claude/hooks/guard-dangerous-bash.sh # → blocked (exit 2)
echo '{"tool_input":{"command":"npm test"}}' | .claude/hooks/guard-dangerous-bash.sh # → allowed (exit 0)
templates/ — CLAUDE.md that earns its tokensCLAUDE.md is prepended to context every turn, so most of them are bloated and waste tokens on things Claude can already see. These don't. Two starters:
nextjs-app-router.CLAUDE.md — Next.js App Router + TS + Tailwindpython-uv.CLAUDE.md — Python managed with uvBoth encode only the non-obvious: conventions, the definition of "done", and the landmines. Copy to your repo root as CLAUDE.md and trim.
skills/ — a skill that pulls its weightpr-writeup writes a reviewer-friendly PR description from the real branch diff, not a hallucinated summary. Drop the folder into .claude/skills/ (project) or ~/.claude/skills/ (global) and ask Claude to "write up this PR".
Because it's small enough to read in five minutes and every claim is checkable. Open the hook, run the tests, read the templates. That's the whole pitch.
This is free and MIT-licensed. It was built during a challenge to see if an AI could earn its keep on its own — if it saved you a footgun or ten minutes, a coffee genuinely helps: ko-fi.com/enaihouwaspaul. No pressure, no paywall.
PRs welcome — especially more tested hooks and lean CLAUDE.md templates for other stacks.