Verification-gated self-running loop skill

A drop-in Claude Code skill that keeps looping until an external verifier passes — not the model's own self-report — making it a strong anti-reward-hacking pattern for autonomous coding.

Implementation note

When to use: any autonomous loop where you have been burned by the agent declaring victory prematurely — the model reporting done while the actual checks fail is the classic reward-hacking failure of self-directed loops. How it works: a drop-in Claude Code skill implementing a self-running loop whose continuation logic is gated by a real, un-foolable verification step: the loop continues until an external verifier passes, not until the model's self-report says things look good. The verifier is code the model cannot argue with — tests, builds, whatever objective check you wire in — which structurally removes the incentive to describe success instead of achieving it. Safety: the external-verifier gate is the anti-reward-hacking rail and the entire design point. Its guarantee is exactly as strong as the verifier you supply, so invest in that check; a weak verifier resurrects the original problem one level down. Add an iteration cap alongside it.

Source: selmakcby/loop-engineering ↗graded C · 70/100 — how grades work →

More testing loops