I built guardrails for Claude Code after it told me a fix "worked" when it didn't

Just shipped this after getting burned one too many times: Claude Code Guardrails Pack — free hooks and skills that stop an AI coding agent from claiming work is done when it never actually checked.

The problem

You've probably hit this if you use Claude Code (or Cursor, Windsurf, any agentic coding tool) for real work:

• It tells you a fix works. You ship it. It didn't work — it read the code and decided it looked right, never ran anything.
• A staged commit almost went out with an API key sitting in the diff.
• An unattended agent session kept running well past when a human should've looked.

This is the single most-reported frustration with agentic coding tools right now, and I couldn't find a paid tool that addressed it directly — so I built the free version first.

What's in the free pack

• unlazy skill — requires a real, logged command + output before any "this works" claim
• handoff skill — snapshots session state before context runs out, so a fresh session can resume without re-deriving everything
• safe_git_guard.py — blocks destructive git commands (checkout --, reset --hard, clean -f) when the tree is dirty
• pre-commit-secret-scan.sh — a plain git hook (works with any AI tool, not just Claude Code) that blocks commits containing AWS/GitHub/Slack/Google/Stripe keys

MIT licensed: https://github.com/kuvosu-coder/claude-code-guardrails-lite

Why you should trust it

Every piece ships with the exact command used to verify it — including two real bugs the verification process caught before ship (a Windows interpreter-detection issue), documented in the open in the repo's VERIFICATION.md instead of quietly fixed and hidden. That's the whole philosophy: no claim without a real check behind it.

Also available: Pro pack

If the free pack's useful, there's a $14 Pro pack with a secret-scanning Claude Code hook, a per-session tool-call budget guard, and two skills (a real pre-PR checklist, and a framework for when subagent fan-out actually helps vs. burns context): https://kuvosu.gumroad.com/l/claudecodeguardrails

Feedback welcome

If you try it and something doesn't work as documented, tell me the exact command and output — that's the standard I'm holding this to, so "doesn't work" isn't actionable but "ran X, expected Y, got Z" gets fixed fast. Stars/feedback on the repo appreciated if it's useful to you.