10/21/2026
Trust, but Break It on Purpose
How do you trust a safety check your AI wrote?
Break the thing it guards, on purpose, and see whether the check notices. A check you have watched catch a real problem is evidence. A check that has only ever passed proves nothing, because it might be working or it might be blind.
Green checkmarks are cheap. Every AI-built system I run is full of them: build checks, broken-link checks, a gate before anything publishes, limits on how slow a page is allowed to be. They're what lets me direct the work without watching every screen. They only help if I know they can fail.
The break-it drill
- Name the faultWhat is this check supposed to catch? A dead link, a missing price, a slow page.
- Cause it on a copyAdd the fault somewhere harmless. Never on the live thing.
- Watch the checkIt should fail, and the message should tell you what broke.
- Undo itRemove the fault and confirm the check passes again.
Ten minutes per check, done once, and the check has earned its place.
Which checks to break first
Start with the checks guarding the things you can't take back or that customers see: anything about money, what's on a public page, and anything that can take a live service down. A check on a throwaway report can wait.
If the check stays green when you break the thing it guards, you've learned something valuable cheaply. Ask your AI to fix the check, then run the same drill again until it fails when it should.
Here's an example of what the drill looks like. Say your AI writes a check that every page has a title. Add a page with no title to a copy of your site. If the check still passes, it was never looking.
This is in addition to review, never instead of it
Breaking a check is one way to trust it. It doesn't replace someone reading what the AI built. Review the work as thoroughly as your goals and time allow, and treat these drills as a second layer. The first layer is knowing which calls stay yours, which I covered in Which Decisions to Hand AI, and Which to Keep.
Part of why I can run several sites by conversation, which I described in Will We Even Need Admin Panels?, is that the checks run before anything reaches me. Each one earns that place by having been broken first.
FAQ
Do I need to be technical to break a check on purpose?
No. Ask your AI to set up the fault on a copy and to show you what the check says. You judge whether the message makes sense to a person.
How often should I repeat the drill?
Once when the check is written, and again whenever the check or the thing it guards changes.
Is a passing check ever enough?
Only once you've seen it fail at least once for the right reason.
Let my Claude help your Claude
Sorting your own work into script, AI and your call is exactly what my one-on-one AI guidance sessions are for: we do it with the systems you actually use, in your business.
Researched and drafted with AI, then reviewed and edited by the site's editors.