The story
I had asked Claude Code to tidy up the audit system on my server — a collection of scripts that run several hundred checks to ensure the server is configured correctly and securely. It came back with a reasonable plan, carried it out, and reported that everything was in order.
I was still getting alerts I didn’t expect. Something wasn’t quite right.
On a hunch, I opened a second tool — GPT — and asked it to look at the same audit system independently. It produced a detailed report. It found documentation drift, overlapping scripts, and two bugs that had been quietly introduced by the very cleanup I had just done. A phantom scheduled job that pointed to a script that no longer existed. And a file whose permissions had been changed in a way that meant the audit script could no longer read it.
Neither of these would have announced themselves dramatically. They would have sat there, generating occasional odd alerts, easy to dismiss as flukes.
The point
[to develop — independent verification is not a new idea. What is relatively new is the ability to apply it to AI-generated code cheaply and quickly. The question worth asking is not whether this is worth doing, but which of your tasks are important enough to deserve it.]