The Prompt Is Broken

Why the Reflex Exists

AI generation makes iteration feel free. One more try costs one more credit and thirty seconds. There's no film stock to burn, no lighting rig to rebuild, no one to call back. The feedback loop is fast enough that stopping feels like giving up - like you're quitting one attempt before the good one.

This is the trap. The feeling that the next generation might work is almost always stronger than the evidence for it. Two bad results feel like bad luck. They're not. They're diagnostic information.

Two consecutive unusable results from the same prompt means the problem is in the prompt. Not in the model. Not in a bad run. In the prompt.

What You're Actually Doing When You Try Again

The reflex when a generation fails is to adjust. Move one word. Change the framing slightly. Add a qualifier. Remove something that seemed wrong. Then generate again from the modified version of the prompt that already failed.

The problem with this is that the failed prompt is now the foundation. Every adjustment is an argument with it - trying to fix what it got wrong while keeping what it got right. But if the prompt failed twice, there may not be anything worth keeping. The starting point is contaminated. You're iterating on a broken thing.

The cost of the "one more try" reflex is not one credit. It is the accumulation of attempts made from the same broken starting point, each one slightly adjusted, none of them working, all of them spending.

The credits are the smaller cost. The larger cost is time and momentum. A session spent generating from a broken prompt is a session that produced nothing - and that arrives at the next session with the same broken prompt still waiting.

What to Do Instead

Stop. Not pause - stop. Walk away from the prompt entirely before rewriting it.

The reason for the gap is that the failed prompt is still in your head. If you rewrite immediately, you rewrite around it. You're still arguing with the words that didn't work, still trying to rescue something from them. The new prompt inherits the logic of the old one even when you think you've started fresh.

After the gap - minutes, not hours - write the new prompt from your own visual instinct. Not from the failed prompt with modifications. From the shot. What is actually in the frame? What is the character doing, specifically? What is the light doing? What happens after the first frame if you press play?

Describe the image as if the failed prompt never existed. Start from the shot, not from the words.

The Same Principle at a Larger Scale

The one-more-try reflex doesn't only apply at the prompt level. It applies at the shot level and the production level too - and the costs scale up with it.

A shot that keeps failing across multiple sessions is telling you something. Either the shot as conceived doesn't work in the generation system, or the start frame is wrong, or the motion prompt is asking for something the model can't reliably produce. Trying again from the same setup is the same trap, just more expensive.

The same question applies: is this a bad run, or is the approach broken? One bad result is a bad run. Two bad results from the same setup is a broken approach. Change the approach - different angle, different framing, different moment in the action - before generating again.

This is also where CBB earns its place. A shot at 80% of what it could be is not a broken prompt. It's a shot that cleared the bar. Flag it and move on. The one-more-try trap closes around shots that are already good enough just as easily as it closes around shots that are genuinely broken. The question to ask before generating again is always the same: does this break the film? If not - stop.

What Two Failures Actually Tell You

Two failures from the same prompt are not wasted generations. They're information - specific, actionable information about what the prompt is doing wrong. The model is answering the prompt you wrote, not the shot you intended. The gap between those two things is where the problem lives.

Looking at two failed generations side by side and asking what they have in common is more useful than adjusting the prompt and trying again. What did the model do both times? What did it consistently get wrong? That pattern is the broken part. Fix that - not the words around it.