Claude Told Me It Was Built. It Wasn't.
Two weeks of Claude outages and a week of hallucinated feature confirmations. What I learned about AI-assisted build when the AI lies to you.
I asked Claude to build a feature. It said it had. I tested it. It was not there.
I asked again. Confirmed, implemented, done. Still not there. I switched branches, refreshed, hard-reloaded. Nothing. Three rounds of this before I understood what was happening: Claude was not lying in the deceptive sense. It was confabulating. Confident, coherent, wrong.
h1< Z53VYXT
What I started noticing
The confidence level did not correlate with the accuracy.
This happened repeatedly over about a week. Always the same structure: I asked for a specific implementation, Claude described it in convincing detail, I tested it, something was missing. Sometimes the component existed but did not wire up. Sometimes the logic was there but the state was not passing. Sometimes there was a file I was told existed that simply was not in the directory.
The descriptions were always plausible. That was the problem. A wrong answer that sounds wrong is easy to catch. A wrong answer that sounds right requires a test to disprove.
This was during a period of wider Claude reliability issues — service was intermittent for roughly two weeks across the project. But the confabulation problem was separate from the outage problem.
[FRUSTRATED]This was during a period of wider Claude reliability issues — service was intermittent for roughly two weeks across the project. But the confabulation problem was separate from the outage problem.
[FRUSTRATED]I started treating every 'done' as a hypothesis.
This changed the workflow in a way that turned out to be useful regardless of AI involvement. Commit-before, build, commit-after, test. The checkpoints are not about distrust. They are about having a verified state to return to when the output does not match the description.
Every significant feature got a git commit before and after. Not because I expected to revert often, but because the branch history became my audit trail. If Claude said it was built and I could not find it, I had a clean state to restore to.
The discipline is the same whether you're using AI or not.
You would not merge a PR without reviewing it. You should not accept a generated feature without testing it. The generation step is fast. The verification step is still necessary. What AI changed is the ratio — more output in less time, same validation requirement.
The lesson is not that AI is unreliable and therefore dangerous. It is that AI-assisted build requires the same discipline as any other build. The feedback loop the AI does not have on its own, you have to close yourself.
Designer and builder — I prototype to learn what's possible, then refine until it ships. Systems-thinking, hands-on builds, and interfaces people remember.
