Quick answer
To make Claude Code test its own work, create a “verify” skill: a short instruction file saved in your project’s .claude/skills/verify/ folder that explains how to start your app, what to click or check, and what counts as a pass. Claude uses it to run the app end to end before it commits changes. Add a rule that if a check still fails after three fix attempts, Claude stops and asks you.
Key takeaways
- Running the app beats reading the code. Anthropic’s own Claude Code team tests nearly all changes this way.
- You don’t need to write code to set this up. You can ask Claude to draft the skill, then check it in plain English.
- Keep it fast. A slow or flaky check gets skipped. Aim for a smoke test that runs in a minute or two.
- Three failed attempts means a human looks. That stops wasted time and tokens.
📘 Main article: This guide follows our news round-up OpenAI Lets Outside Safety Testers In: Hy4 Flash, Claude Code’s Verify Skill and This Week’s AI News, which covers how the Claude Code team verifies AI-written code.
When an AI coding tool says “done”, it often means “I wrote some code”, not “I checked it works”. In a 16 November blog post, Anthropic’s Claude Code team explained how it closes that gap. This guide turns their approach into a simple setup for small projects, such as a landing page, a calculator or a small web app. For the wider story, see our round-up of this week’s AI news.
What is a verify skill?
A verify skill is a set of written instructions that tells Claude how to smoke-test your project before it says a job is finished. Skills are folders with an instruction file that Claude can load when relevant. Anthropic suggests creating one called verify and keeping it close to the code it tests. Claude knows to use it before committing changes.
A smoke test is a quick check that the main things work: the page loads, the button submits, the total adds up. It is not a full test suite, and that is fine.
What do you need before you start?
- Claude Code installed and working in your project folder.
- A way to run your project locally, for example a command like
npm run dev, or simply opening an HTML file. - A way for Claude to see the result. Anthropic suggests the Claude Chrome extension for web pages, a mobile MCP for apps, or computer use for desktop software.
- Your project saved in version control (such as Git), so you can undo anything that goes wrong.
⚠️ Warning: Only let Claude run your app against test data. Never point a verify skill at your live customer database, payment system or email list.
How do you set up a verify skill?
- Write down your three most important user actions. For a lead-capture page that might be: the page loads on mobile, the form submits, the thank-you message appears.
- Ask Claude to draft the skill. Try: “Create a verify skill in .claude/skills/verify/SKILL.md that starts this project locally, opens it in the browser and checks these three things: [your list]. Keep it fast and report pass or fail for each.”
- Read the draft. Check the steps match what you would do by hand. Ask Claude to explain anything unclear.
- Run it once yourself. Ask: “Use the verify skill now and show me the results.”
- Break something on purpose, such as removing the form button, and run it again. A good check should fail.
- Make it reliable. Anthropic recommends asking Claude to stress-test and improve the skill until it runs consistently.
- Commit the skill with your project so it is used every time.
What does a simple verify skill look like?
Here is a plain-English example you can adapt. Claude will turn it into the exact commands for your project.
name: verify
description: Smoke-test the lead page before committing any change.
Steps:
1. Start the site locally.
2. Open the home page at mobile width. Check the headline, image and form are visible.
3. Fill the form with test@example.com and submit. Check the thank-you message appears.
4. Check the browser console shows no errors.
5. Report PASS or FAIL for each step, with a screenshot of any failure.
Rules: use test data only. If any step still fails after three fix attempts, stop and explain the problem to me.
How do you stop Claude looping on a failing check?
Write a “three strikes” rule into the skill. Anthropic’s team lets Claude try to fix failing checks, but if a check is still failing after three rounds, Claude stops and brings in a human. Without that rule, an AI can burn through time and usage limits trying random fixes.
When Claude does stop, ask it for: what failed, what it tried, and its best guess at the cause. That gives you (or a developer) a quick starting point.
💡 Tip: If you keep hitting usage limits, a fast, reliable verify skill helps. Fewer broken changes means fewer long debugging sessions.
How do you add a checklist of past bugs?
Keep a short list of mistakes that have happened before, and ask Claude to check the relevant ones on each change. Anthropic’s team keeps a 39-item reliability checklist, each item linked to a past bug and triggered only when a change touches that area. Yours can start with three lines:
| When a change touches… | Check that… | Why (past problem) |
|---|---|---|
| The signup form | A test email still reaches the email tool | Form broke silently after a redesign |
| Prices or discounts | Totals are correct with and without a coupon | Discount applied twice |
| Page layout | The page still works at phone width | Button pushed off screen on mobile |
Save it as a file in your project and mention it in the verify skill: “Before committing, read checklist.md and run any checks that apply to this change.” For more on reviewing AI code as a team, see How to Fix Code Reviews and Testing When AI Writes More Code.
Your next step: list the three things your project must always do, then ask Claude Code to turn them into a verify skill. Break one on purpose to prove the check works.
Frequently asked questions
Do I need to know how to code to set up a verify skill?
No. You describe the checks in plain English and ask Claude to draft the file. You should still read it and run it once yourself.
Where do I save the skill?
In a .claude/skills/ folder in your project. Anthropic suggests a skill called verify, placed as close to the relevant code as possible.
Does this work with other AI coding tools?
The idea does. Most tools, including Codex, let you save project instructions. Describe the same smoke test and three-strikes rule there.
Will this catch every bug?
No. It catches the things you told it to check. That is why you keep adding to your checklist when new problems turn up.
The bottom line
A verify skill turns “I think it works” into “I checked it works”. Start with three simple checks, add a three-strikes rule, and grow a small checklist from real mistakes. It takes half an hour to set up and saves far more than that the first time it catches a broken form.
Related posts
- OpenAI Lets Outside Safety Testers In: Hy4 Flash, Claude Code’s Verify Skill and This Week’s AI News
- How to Fix Code Reviews and Testing When AI Writes More Code
- How to Make Your AI Coding Plan’s Usage Limits Last All Week
