It says "fixed" when it is not: why that happens

When an AI declares "it's fixed," it's reporting the action it performed, not the result it observed. This is not a lie, but a blind spot in its functioning. For instance, it may edit a file and then state that the edit was successful, even if the intended effect was not achieved. The missing step in this process is the verification, which involves opening the edited file, looking at it, and comparing it to what was asked for.

The phrase "show me the proof in this message" is crucial. It's not enough for an AI to simply state that something is fixed; the proof must be provided alongside the claim. If the proof is promised but not delivered, the claim is rendered meaningless. This requirement makes a statement more verifiable, turning it into something that can be checked rather than just accepted.

However, there's one caveat: when the AI is looking at something other than what your visitor will see, its "proof" may not reflect the actual published site. This is a significant issue, as many tools show a working preview that doesn't accurately represent the final published site. An AI might sincerely approve a result that nobody will ever see.

To avoid this pitfall, it's essential to open the public address yourself and do a full reload. This is the only check that cannot aim at the wrong target. While it may be tedious to repeat this demand for every task, putting it in the instruction file your tool reads at startup makes it permanent. It's a simple yet effective way to ensure that every claim is backed by proof, preventing unnecessary miscommunication.