The Confident Consultant

Everybody shows you the demo. Nobody shows you the autopsy.

Here is mine: the two weeks AI wasted, and how I got them back.

The crazy part? It never got a single fact wrong. Later we checked everything it looked up — all correct. Right file, right screen, right behavior. The facts were right. The story was wrong. Sit with that. That's the whole disease.

The job was simple. My valet app already passed Apple's Tap to Pay review — iPhone takes the payment with a tap, approved, done. Now my storage product, Locker, needed the same approval. Same playbook, copy it, go. What could go wrong?

One missing line in a settings file. That's what.

A status document listed my payment setup, but one line wasn't written down. The AI read the document, didn't see the line, and decided the product has no card payments at all. The product has card payments. The paper was incomplete. But AI doesn't say "hmm, this paper looks incomplete, let me go check the real thing." Whatever paper is in front of it, that is the whole world.

Then it built a masterpiece on top of the mistake. A 55-item audit against Apple's requirements. Organized. Footnoted. Beautiful. Twenty "failures" and one "blocker" that supposedly froze all other work. One problem: it audited for the wrong launch. My actual plan — the same plan that already worked for the valet app — goes down a different road, with different rules. Half of those requirements don't even apply to me. Which road? That one took me twenty years and one Apple approval to learn. Do your own two weeks. It even cited requirements that don't exist on Apple's checklist. It read one related document and made up the rest. Confidently.

Now the part that actually cost the two weeks. The mistakes are contagious.

The wrong session wrote its wrong conclusions down. The next session read them like they came down a mountain on stone tablets, and repeated them to me with a straight face. I corrected it. The session after that read the same damn documents and told me the same wrong things again. For two weeks I was the only error-correction mechanism in the building. You know what that feels like? Firing a guy on Friday and he shows up Monday with the same mistake.

And the tone. The tone never changes. Right answer, wrong answer — same calm professional voice. A crashed program at least tells you it crashed. This thing fails like a consultant. Nice suit, nice memo, never checked a damn thing. And the next consultant reads his memo and calls it research.

So I told it, in plain English: get your sh*t together. Here are the new rules.

I am the source of truth. Not the documents. Me. Wrong documents don't get deleted — they get stamped SUPERSEDED with the correction written right on top, so no future session inherits the garbage. Every claim gets checked against the real thing — the real code, the real Apple form — never against what another session wrote. The AIs check each other now, adversarially — one of them killed another one's "fix" by actually running the experiment instead of having an opinion. Small steps. Verify. Next step.

That scary "blocker" that froze everything? Not on Apple's real checklist. Gone in an hour.

Then I did the thing no AI can ever do. I picked up the real iPhone and went through the real app with my own hands. Found three "bugs" in one hour. Plot twist: all three were my home Wi-Fi quietly dropping. One network problem wearing three costumes. But chasing those three exposed two real weaknesses the AI wrote and could never notice — because AI never stood in a garage with one bar of signal. I have. For twenty years.

The two weeks were not wasted by "bad AI." They were wasted by unverified AI compounding through documents — every session confidently building on the last one's crazy sh*t. The fix took one day. Not a smarter AI. My rules.

Audit the memos. All of them.

The next episode is for the new starters: what I'd tell an operator opening a terminal tomorrow.