# The Checker Gate: Your AI Says "Done." Stop Believing It.
Your AI just told you it's done. The tests pass, the fix is in, the deploy is clean — except one of those claims is a lie, and you find out in production. The fix isn't a smarter model. It's a second set of eyes: a checker agent that re-runs every "done" claim against the actual files, with zero loyalty to the one who did the work. The maker never grades its own homework. This page hands you The Checker Gate — the exact verifier agent file, the 4-point checklist it runs, and a worked example of a false-done getting caught.
## Why false-dones happen
An AI that just built something is the worst judge of whether it works. It re-reads its own output with the same assumptions that produced the bug. It says "done" because the code *looks* finished, not because anything verified it. Every developer running agents has shipped at least one of these — and every false-done that reaches production costs 10x what catching it at the gate would have.
The full step-by-step (2 minutes)
1. Drop the agent file in. Copy verifier.md from the kit into ~/.claude/agents/ (create the folder if it doesn't exist). That's the install.
2. Route every "done" through it. When your main session (or any subagent) claims a task is finished, spawn the verifier with one line: what was claimed, and where the evidence should live. It reads the actual files — it never trusts the claim's own prose.
3. It runs the 4-point checklist:
- Does the claimed change actually exist in the files right now?
- Does the proof command (test, build, curl) pass when re-run fresh?
- Do the hard rules hold (nothing broken next to the fix)?
- Is anything claimed that has no evidence at all?
4. PASS ships. FAIL goes back. The verdict comes back with one-line reasons. A FAIL goes back to the same worker with the exact gap — not to you, and not to production.
5. Keep the roles separate, forever. The checker never fixes things itself, and the maker never checks itself. That separation is the whole mechanism — the moment one agent does both jobs, you're back to grading your own homework.
What's in the kit
- verifier.md — the drop-in checker agent file, sanitized and ready for any Claude Code setup
- CHECKLIST.md — the 4-point gate, printable
- **EXAMPLE.md** — a real worked example: three "done" claims, one caught as false before it shipped
Comment CHECKER on today's video and the kit lands in your DMs — or grab it right here with the button above. Fair warning: the comment word rotates out when tomorrow's drop lands at 4pm.