You want Claude Code to keep working on a change until it is genuinely done, rather than stopping at its first plausible attempt. What should you give it?

The principle — an agent finishes when a check passes, or else when it decides it has finished. Only one of those is under your control.

Why the key is correct — Anthropic puts it plainly: give Claude something that produces a pass or fail, and the loop closes on its own. The glossary defines a verification loop as what happens when you give Claude a check it can run, such as a test suite, a build or a screenshot comparison, so that Claude iterates until the check passes instead of stopping after one attempt. The important phrase is a check it can run: a failing suite is information the loop consumes, and it is what turns the third phase of the agentic loop from a formality into a real gate.

Why the others are wrong — telling it to be thorough supplies an adjective rather than a criterion, leaving the standard as its own opinion. A longer description sharpens the first attempt without saying when the work is complete; specification and verification are different problems. And asking it to report back before finishing hands the judgement to you, which stops the loop rather than letting it run until something objective passes.

Remember this — before delegating, ask what you would run to know it worked, and give that to Claude.

Sources — Anthropic's Best practices for Claude Code and the Claude Code glossary.