An experiment protocol that fits on a card

Why most prompt advice is wrong

Somebody changes four words, gets a better clip, and posts the new prompt as a discovery. The clip was better because the random draw was better. The four words did nothing.

This is not carelessness by unusual people. It is the default outcome of how these tools present themselves: one button, one result, no controls. You have to impose the discipline yourself, and almost nobody does, which is why the public prompt literature for video is largely folklore.

The protocol

Five rules. They fit on a card and they are worth more than any prompt guide.

1. Fix the seed. If your tool exposes one, set it. If it does not, generate four takes of each condition and compare the sets rather than the individuals.
2. Fix the still. Same first frame for every condition. A different still is a different experiment.
3. Change exactly one thing. One word, one slider, one setting. Not a rewritten sentence.
4. Write it down before you look. What you changed, and what you expect. Recording the expectation is what stops you from reinterpreting the result afterwards.
5. Compare against the previous file, not your memory. Put them side by side. Memory of a clip you watched four minutes ago is unreliable in exactly the direction of whatever you hoped for.

A worked example

Suppose you want to know whether naming a focal length does anything on your model.

• Still: one frame, fixed. Seed: 42, fixed. Everything else: fixed.
• Condition A: medium shot, a man at a workbench, static camera
• Condition B: the same plus shot on an 85mm lens
• Condition C: the same plus shot on a 24mm lens

Three generations. Then look specifically at the relationship between the subject and the background: does B compress it and C stretch it? If both look identical to A, your model does not respond to focal length and you have just deleted a category of words from your prompts permanently.

That is a real finding, it took fifteen minutes, and it is about your model rather than somebody else’s.

What to test first

Your free allowance is finite, so spend it in order of how much the answer changes your work:

1. Motion strength. Find the value where artefacts start. This affects every clip you will ever make.
2. Does the model reframe from the still? Ask for a close-up from a wide still. If it obeys, size belongs in the prompt. If it does not, size belongs in the frame.
3. Camera move vocabulary. Six phrasings, keep the ones that produce a distinct correct move.
4. Guidance. Three values. Watch motion die at the top.
5. Does the negative prompt do anything at all?
6. Step count. Find where improvement stops.

Six experiments, somewhere between twenty and forty generations, and you will know your model better than most people using it professionally.

Keeping the record

A plain text file per model:

MODEL: <name and version>, tested 2026-09
- motion: artefacts begin at 7/10; production default 4
- reframing: ignores requested shot size, obeys the still
- moves that work: slow dolly in / handheld follow / static locked-off
- moves that fail: crane up (becomes a tilt), whip pan (smears)
- guidance: 5 good, 8 freezes motion
- negative prompt: no measurable effect (distilled)
- steps: no visible gain past 22

Seven lines. It will save you money every week and it becomes worthless the moment the provider updates the model, which is why it carries a date and why you re-run it after an announced update.

The trap of the impressive result

One last warning. When an experiment produces a beautiful clip, the temptation is to stop and keep the prompt. Resist it for one more generation: run the same condition with a different seed. If the second one is also good, the change did something. If it is ordinary, you found a good sample rather than a good instruction.

This single extra generation is the difference between building knowledge and collecting superstitions.

Today: run experiment one from the list (motion strength) and write the resulting number at the top of your notes file. Use it as your default from now on.