The third try was a different idea

When an AI tool misses, the instinct is to explain more. Often, that's exactly what makes the next version worse.

Every detail you add is one more thing it has to get right in the same pass. Ask for an email that's warm, firm, mentions the deadline, apologizes for the delay and doesn't sound like a template, and one of those can slip. Add a sentence to fix it, and something else slips. The request gets longer and the misses just move around.

Splitting the request works better: ask for one thing, get it right, then build on it. So does looking at what you already have before you ask again. A version you rejected often has a piece that's fine, just in the wrong place.

I learned both on a few seconds of video for my show. A buyer is meant to walk away from a deal. Nothing is filmed; a model generates the shot from a still and a written prompt. The first version came back too fast. So I asked for more: push the chair back, stand, walk off shaking his head. The second missed the chair push and the head shake, and looked rude. In my takes, the model was more reliable with one action per shot than a chain of them. Asking for more was making it worse.

So I split it: a handshake, then a hard cut to him walking away. The fix reused the walk from the first version, the take I'd sent back as too fast. After a handshake, it isn't too fast.

That scene took 43 minutes from first note to lock. The whole episode took 79 rounds of review to lock. Next time a take misses on three things, I'll ask for one of them, and look at what I sent back before I write another word.