What actually changes a result, and what does not

We ran eight scenes through every model in the catalogue, five seeds each, changing one thing at a time. The result was duller than we hoped and more useful than we expected: most of what people put in a prompt does nothing at all, and a short, boring list of things does nearly all of the work.

Here is that list, in the order the changes actually showed up on screen, and then the things we stopped typing.

The six that move a render

Ordered by effect size, where effect size means: how often a rewrite that touched only this one thing produced a visibly different result across five seeds.

• Subject and action, in that order. A woman closes a laptop is a shot. A woman, laptop, evening, moody is a mood board, and a video model asked for a mood board will invent an action for you.
• Shot size. Close-up, medium, wide. This is the highest-leverage word in a video prompt, because it decides what the model has to keep stable for the length of the clip. A wide shot has to hold a whole body upright for eight seconds. A close-up only has to hold a face, and it holds it far better.
• Camera move. Static, slow push in, handheld follow, orbit left. Name one and you get one. Name none and you get whatever the model reaches for, which on both video models is a slow push in roughly two thirds of the time.
• Light, named by source and direction. Late afternoon sun through a window, camera left beats warm cinematic lighting every time. The first is a physical fact a renderer can resolve; the second is an adjective about how you want to feel afterwards.
• Material and surface. Wet asphalt, brushed steel, raw linen, chipped enamel. Materials are how light becomes visible, so naming one is how you get texture without asking for texture.
• Speed. Slow and deliberate, or quick and clipped. Video models default to a mid-tempo drift, and if you want stillness or urgency you have to say so — it is not going to be inferred from the subject.

The four that never have

Each of these was added to, and removed from, the same prompt across the same seeds. None of them moved anything we could point at.

• Quality adjectives. Cinematic, masterpiece, highly detailed, award winning. This was real prompt-engineering craft two model generations ago on a very different training distribution. It is folklore now. Adding all four at once changed nothing measurable on any model we run.
• Resolution words. 4K, 8K, ultra HD. Resolution is a parameter on the generation, not a description of the scene. Typing it into the prompt spends your attention and buys nothing.
• Negations. No blur, without distortion, not cartoonish. There is no reliable mechanism for this in prompt text on these models. In practice no blur is read as an instruction containing the word blur, and you get more of it, not less.
• Camera bodies and film stock. Shot on ARRI Alexa, Kodak Portra 400. These do sometimes shift the colour a little, but never in the direction printed on the box. If you want warm and slightly green, ask for warm and slightly green — you will get it more often and you will know why.

The same prompt, both ways

This is one of the eight scenes, before and after. The second version is shorter, and it is the one that produced the same shot four times out of five.

Before
  cinematic 4k masterpiece shot of a chef in a kitchen,
  highly detailed, dramatic lighting, no blur, trending

After
  Medium shot, static camera. A chef lifts the lid off a
  steel pot and steam rises into hard morning light from a
  window, camera left. Wet steel, dark tiles. Slow.

The after version is not more poetic. It is more specific about six things and silent about everything else, which is the whole technique.

Where this changes by model

The six hold everywhere, but they do not hold equally. Omni Flash follows a camera move instruction more literally than Veo 3.1 Lite does, and it will happily hold a static frame when you ask for one. Veo 3.1 Lite is better at keeping a face the same face from the first frame to the last, and it is more willing to overrule you on framing if it thinks the shot needs the room.

For stills, Nano Banana Pro responds to material and light far more than to shot size, which makes sense — there is no time axis to keep coherent, so the model spends its effort on surfaces.

How to use this tomorrow

• Write the shot before you write the mood. Subject, action, shot size, camera, light. If the sentence still needs a mood word after that, keep it. Usually it does not.
• Delete every adjective that describes the output rather than the scene. That one rule removes about a third of most prompts and costs nothing.
• Change one thing per re-run. If you rewrite the whole prompt after a bad result you learn nothing, and you will do it again next week.
• Keep the ones that worked. Any published result in Explore carries the exact prompt that made it, and you can copy it, edit one clause, and run it.

None of this is a trick and none of it is model-specific enough to go stale. It is closer to writing a shot list than to programming, which is the good news: the skill transfers, and it transfers to the next model too.