Give Claude a way to check its own work

Claude stops when the work looks done. Without a check, that is the only signal it has.

You become the person who has to spot every mistake.

Why this matters

When you ask Claude to do something, it runs until it thinks the job is finished. But "looks done" and "is done" are not the same thing.

A spreadsheet might have a hundred rows. Claude might fill them all and stop. But if row seventy-three is missing an email address, you have to find it.

A check gives Claude a way to know whether the work passed or failed.

How it works

A check is anything that gives a clear pass or fail.

It could be a test that runs code and reports errors. It could be a build that either completes or breaks. It could be a screenshot compared with a design.

For non-coders, a check is often simpler. You tell Claude what "done" looks like before it starts. Then you ask it to prove it at the end.

For example, you might say every row must have an email address and no duplicates. At the end, Claude runs through the file and reports what it found.

Anthropic's guide recommends asking Claude to show evidence rather than just saying it succeeded. That means the command it ran and what came back, or a screenshot, or the test output.

The guide also suggests a second opinion. A separate reviewer, with a fresh view, checks the work against the original plan and reports gaps.

It warns that a reviewer asked to find gaps will usually find some. So ask it to flag only gaps that affect correctness or the stated requirements.

How to do it

1. Say what "done" looks like before Claude starts. Be specific.
2. Ask Claude to check the work at the end and show you the evidence.
3. If the task is complicated, ask a second reviewer to check the work against the plan and flag gaps that matter.

Here is an example prompt for a simple task:

> I have a spreadsheet of customer names and email addresses. Some rows are missing emails and some emails appear twice. Clean it up so every row has an email address and there are no duplicates. At the end, count the rows and check that every email address contains an @ symbol and appears only once. Show me the count and any problems you find.

And here is a prompt for a second reviewer:

> Look at the work Claude just did. Check it against the original plan. Flag any gaps that affect correctness or the stated requirements. Do not flag style issues or minor improvements.

Common mistakes

Asking Claude to check its own work without saying what "done" looks like. Claude will check something, but it might not be what you need.

Accepting "I have finished the task" as proof. Ask for evidence instead.

Asking a reviewer to find every possible gap. It will find many. Ask it to flag only gaps that matter.

One thing to try this week

Pick a task you do often. Next time you ask Claude to do it, add a check at the end that confirms the work meets your requirement.

For example, if you ask Claude to tidy a list, tell it to count the entries and confirm none are duplicates. Ask it to show the count and report any duplicates it finds.

You might be surprised what it catches.

The free What could your business automate? check takes about two minutes.