Guides

How to check AI-written code when you're not a developer

The short answer: You don't have to read the code to check it. Check what the app does, make Claude show proof that it checked, ask a fresh reviewer, and bring in an expert for anything involving money, sign-in or personal data.

Key points

  • Give Claude a check that comes back pass or fail, so it can prove its work instead of stopping when it looks done.
  • Ask for evidence, not reassurance: the exact command it ran, and its output.
  • Have someone who did not write the code review it, and tell them to skip style.
  • Save your work in Git after every working change, so a bad change costs minutes.
On this page
  1. Can I trust code that Claude wrote?
  2. What can I check without reading code?
  3. How do I make Claude check its own work?
  4. What should I ask Claude to show me?
  5. Can Claude explain code I can’t read?
  6. Who should review the code?
  7. How do I stay able to undo things?
  8. When do I need an expert?
  9. What does a good routine look like?
  10. A short checklist

Can I trust code that Claude wrote?

Trust it the way you would trust work from a fast new colleague: it is often right, and you still check what matters. Claude can say “done” when something is only almost done. The best-practices page explains why: Claude stops when the work looks done, and if it has nothing to run that tells it whether the work really is done, “looks done” is all it has (Claude Code docs: best practices).

The same page says that without such a check, you become the test suite, and every mistake sits there until you happen to notice it. The way out is to build checks, and you can do that without reading code.

What can I check without reading code?

You can check behaviour: what the app does when you use it. This is what a client or a customer will do anyway.

  1. Walk the main path. Sign up, do the main thing, come back tomorrow. Does it work?
  2. Break it on purpose. Leave a field empty. Type something silly. Press the button twice. Go back mid-way. Switch off the internet.
  3. Try it as someone else. Make a second account. Can it see the first account’s data? It must not.
  4. Check it on a phone. And on a small screen. Many apps are only tried on one laptop.
  5. Look at what it says when it fails. A clear message is a good sign. A blank screen is not.

Write the walk-through down as a short list. It becomes your checklist for every change.

How do I make Claude check its own work?

Give it something that comes back pass or fail. The best-practices page lists what counts: a test file, a build’s exit code, a linter, a script that compares output with an expected result, or a browser screenshot placed next to a mockup.

You don’t need to write these. Ask Claude:

Ask Claude

Add automated tests for the main path of this app, and tell me which important cases they do not cover.

Then put the command that runs them into your CLAUDE.md. The docs call that command the single most useful line in the file if your project already has a check.

You can also say it in the prompt: run the tests after the change and keep going until they pass. The docs note that this works for any task today, with no setup.

What should I ask Claude to show me?

Ask for evidence, not reassurance. The docs say: “all tests pass” on its own tells you very little, and reading the output is faster than running the check yourself again.

Good things to ask for:

  • The exact command it ran, and its output.
  • A screenshot of the result, if it is something you can see.
  • A list of what it did not check.
  • A plain explanation of what changed, as if explaining to someone who doesn’t code.

Tip

Once Claude says it’s done, the docs also suggest running /verify in the session, which has Claude build and run the app and look at the result. The docs say this catches the kind of problem that tests and type checks miss.

Can Claude explain code I can’t read?

Yes, and this is one of the best uses. Ask:

  • “Explain what this change does in three sentences, for someone who doesn’t code.”
  • “What could go wrong with this change?”
  • “Which files did you change and why?”
  • “What did you assume that I didn’t tell you?”

The last question is the most useful. A lot of bad surprises are assumptions that nobody said out loud.

Who should review the code?

Someone who did not write it. The best-practices page describes /code-review, which reviews your changes in a fresh subagent that never saw your conversation, so it isn’t swayed by the reasoning that produced the code.

Watch out

If you ask for a review, tell it what to look for. The same page warns that a reviewer asked to find problems will always find something, so ask for problems that affect correctness and tell it to skip style. Otherwise you end up with extra layers and tests for cases that can’t happen.

How do I stay able to undo things?

Always be one step from a working version.

  • Claude Code keeps checkpoints of the files it edits. /rewind (or Esc twice on an empty prompt) lets you roll code, conversation or both back (checkpointing docs).
  • The docs are clear about the limits: files changed only by commands aren’t restored, and checkpoints work alongside version control rather than replacing it.
  • So save your work in Git after every working change. Ask Claude to commit with a clear message.

With a working copy always saved, a bad change costs you a few minutes instead of a weekend.

When do I need an expert?

Check this list. If any of these are true, have a person who reads code look at it before real users arrive:

  • It takes payments or handles money.
  • It has sign-in, or users can see each other’s data.
  • It stores personal information such as emails, addresses, health or messages.
  • It connects to someone else’s system with a key or password.
  • A mistake would be expensive or embarrassing.
  • You can’t tell what it does and you can’t make Claude explain it clearly.

The Claude Code security page puts the responsibility in one line: you’re responsible for reviewing proposed code and commands for safety before approval (Claude Code docs: security). If you can’t review it yourself, getting someone who can is how you meet that responsibility.

An expert review is not a sign that you failed. It is a normal step for anything that matters, and it goes faster when you hand over a clear CLAUDE.md, a plan and tests.

What does a good routine look like?

For each change:

  1. Say what you want and what “done” looks like.
  2. Let Claude build it and run the tests.
  3. Ask for the evidence.
  4. Walk the main path yourself, and try to break it.
  5. Ask for a fresh review of the change.
  6. Save it in Git.

It sounds like a lot, and it takes minutes once it is a habit.

A short checklist

  • The main path works, as you use it, on a phone and a laptop.
  • You tried to break it.
  • Another account can’t see the first account’s data.
  • Tests exist, and Claude showed you their output.
  • A fresh reviewer looked at the change.
  • The work is saved in Git.
  • Anything with money, sign-in or personal data has been seen by an expert.