Iterate an image prompt based on exactly what went wrong last time
Regenerating blindly wastes attempts. Diagnose the miss, then fix that specifically.
- Use it for
- Anyone generating images who keeps getting close-but-not-quite results.
A common failure loop with image generation is regenerating with vague adjustments ("make it better," "try again but nicer") that don't target what actually went wrong, burning generations without converging on the result you pictured. The fix is treating each attempt as diagnostic information: what specifically differs between what you got and what you wanted, and which part of the prompt is responsible for that gap.
This prompt structures that diagnosis explicitly instead of guessing at the next attempt.
When not to use this
If the first result was already close and you just want minor variation, this level of diagnosis is overkill — just ask for a variant. This is for when you're several attempts in and still missing the mark in a way you can't quite name.
Did this work?
The next generation fixes the specific thing that was wrong last time, instead of changing randomly and hoping — you can point to which prompt change targeted which miss.
Tested on claude-opus-5. Evidence status is draft; it moves to battle-tested only on recorded runs, never by hand.