Claude Code Guardrails: stop false "done" reports
Three bash hooks that surface failed commands, force checked-vs-unchecked replies, and break retry loops.
Free
Works with the AI tools you already use
Claude Code
Cursor
Codex CLI
GitHub Copilot
Gemini CLI+20 moreClaude CodeCursorCodex CLIGitHub CopilotGemini CLI
OpenClaw
Manus
Grok+17 agentsClaude Code Guardrails: stop false "done" reports
Example session with this skill installed
Install the guardrail hooks for this project. Make sure we have the loop breaker and the verification quantifier.
- Read your context and instructions
- Compiled the claude code guardrails
Installed the three guardrail hooks into this project. This is the real output from bash install.sh ~/myproj (macOS, jq 1.7):
installed ~/myproj/.claude/hooks/verify-before-done.sh
installed ~/myproj/.claude/hooks/quantify-claims.sh
installed ~/myproj/.claude/hooks/loop-breaker.sh
settings ~/myproj/.claude/settings.json (backup written alongside)
json ok
Restart Claude Code, or open /hooks once, to load them.
Check
| What | Command | Result |
|---|---|---|
| Hook events registered | jq '.hooks | keys' .claude/settings.json | ["PostToolUse", "UserPromptSubmit"] |
| Backup written | ls .claude/ | settings.json.bak.20260929125322 |
| Second run | bash install.sh ~/myproj | adds no duplicate entries |
What the agent sees after a failing test run
Fed a Bash result of === 2 failed, 14 passed === with exit code 1, verify-before-done.sh returned this context to the agent:
The command that just ran shows failure signals: failing tests.
Do NOT report this turn as done, fixed, working, or passing until you have
addressed it. Required:
1. Read the actual output and exit code. Do not infer them.
2. State plainly what failed and why.
3. Anything you could not verify, label as unverified.
Next steps
- Restart Claude Code or open
/hooksonce so the hooks load. - To remove: restore the newest
settings.json.bak.*, or delete the three entries whose command ends inverify-before-done.sh,quantify-claims.shorloop-breaker.sh.
Connects securely to your tools. The creator never sees your data.
About this skill
Claude Code sometimes summarizes a turn as finished when a command in it failed. A rule in CLAUDE.md does not reliably prevent that: in one workspace the verification rule was in context for all 22 verification-shaped replies and was followed in 13 of them (59%). Hooks fire at the moment the rule matters, so these three move the rules out of the instructions file.
What is included (3 bash hooks, 58-69 lines each, bash + jq only):
- verify-before-done: after each Bash call, scans the output for failure signals (non-zero exit, failing test counts, tracebacks, connection refused, permission denied) and adds a note naming what it found. Skips cat/grep/head/--help so reading a file about failures does not fire it.
- quantify-claims: on verification-shaped prompts, asks the reply to open with "verified n% (numerator/denominator)" and to name what was not checked.
- loop-breaker: counts consecutive calls to the same tool and, past a threshold, tells the agent to stop and choose a different move. Ordinary tools (Bash, Read, Grep, Edit, Write) are skipped by default.
install.sh merges into your existing settings.json, writes a timestamped backup first, and adds nothing on a second run. Nothing is sent anywhere; the hooks only read the hook payload on stdin. Read the hooks before installing.
What it does not do: it does not make an agent honest or prevent every false completion. It makes failed output hard to skip past.
How to install
Works the same in every agent - Claude, Cursor, Codex, Copilot and 20+ more.
- 1
Download the ZIP
Free skills download straight away. Paid skills unlock right after purchase.
- 2
Unzip into your skills folder
Every agent reads skills from one folder on your machine. Drop the unzipped folder in there.
- 3
Ask your agent to use it
Restart the agent if it was already running. It picks the skill up automatically - no config needed.
Skills folder by agent
Click the path to copy it. Create the folder if it does not exist yet.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scanned
Verified clean 2 days ago
- Free to download with an account