Change One Thing: How To Iterate On A Prompt

The usual way people improve a prompt is to look at a disappointing result, rewrite most of the prompt, and run it again. Then repeat, six or seven times, until something acceptable appears.

At the end of that you have an image you can use and no idea why it worked. The next shot starts from zero.

Change one clause at a time

The alternative is dull and much faster in aggregate: keep the prompt fixed and change exactly one thing per run.

If the framing is wrong, change only the camera clause. If the light is flat, change only the lighting clause. Resist the urge to also fix the two other things you noticed, because if the next result is better you won't know which change did it.

This works because the six slots — subject, action, setting, camera, light, style — are largely independent. You can swap the lighting clause without touching the subject and the subject will mostly survive. That independence is what makes single-variable iteration practical.

Keep the version that worked

Save prompts that produced good results, verbatim, with the model and settings they ran on. Not a description of them — the actual string.

This sounds like bookkeeping and it is, but the alternative is regenerating from memory and getting something subtly different. A prompt that reliably produces a particular look is a reusable asset; the same prompt half-remembered is not.

Worth storing alongside it: the model, the aspect ratio, the resolution, and the seed if the model exposes one. Any of those changing will change the output, and "the prompt stopped working" is nearly always one of them moving rather than the prompt.

Use a seed when you have one

If the model accepts a seed, fix it while you iterate.

A seed pins the randomness. With it fixed, a change in output is caused by your edit rather than by chance — which is the entire point of single-variable iteration and impossible without it. Change the seed only when you want a different roll of the same prompt.

Not every model exposes one. Where it isn't available, run the baseline twice before concluding your edit did something: if two runs of the unchanged prompt already differ noticeably, that variance is your noise floor and small improvements aren't measurable below it.

Know when the words aren't the problem

There's a point where rewording stops helping, and it comes sooner than most people accept. Signs you've reached it:

• The same failure appears across five or six substantially different phrasings.
• The model produces the thing you asked for but badly — mangled hands, illegible text, physically impossible geometry.
• You're adding qualifiers ("actually", "make sure", "it is important that") — which is arguing with a system that isn't listening.

At that point the fix is a different input, not a better sentence:

• Give it an image. A reference or first frame settles composition and subject appearance in a way no amount of description will.
• Split the work. Generate the still, fix the one broken element in an edit pass, then animate. Asking one generation to get everything right at once is the actual problem.
• Change the model. Text rendering, hands and faces vary enormously between models, and a weakness in one is not something you can prompt around.

Iterate cheap, finalise expensive

Do all of this at the lowest resolution and shortest duration the model offers. Composition, framing and lighting are all fully legible at 480p. Once the prompt is right, run it once at the settings you actually need.

The common mistake is iterating at final quality, which makes every attempt cost several times what it needs to and quietly discourages the extra attempts that would have got you there.

A working loop

1. Write the full six-slot prompt.
2. Generate at low resolution, short duration, fixed seed if available.
3. Name the single worst thing about the result.
4. Change only the clause responsible.
5. Repeat from 2 until nothing is clearly wrong.
6. Save the prompt with its model and settings.
7. Re-run once at final resolution and duration.

Step 3 is the one people skip, and it's the one that makes the rest work. "It doesn't look right" is not actionable. "The light is flat and frontal" tells you exactly which clause to edit.