The anatomy of an AI prompt: what each added line really changes

I had the story in my head before I made the pictures. You type six words, the model gives you a three-legged player, and you learn you should have written a better prompt. It's the story you read everywhere. I only wanted to illustrate it.

The model didn't play along. What it showed me was less dramatic, and more useful. So here is what I actually got, with the prompt, the seed and the time under every picture.

A prompt is an order you give a cook

Before the pictures, a mental picture. When you order "a plate of pasta", the cook decides everything else: the sauce, the plate, the portion. They fill in the blanks. An image model does the same. Whatever you don't say, it chooses for you.

The seed is where the randomness starts. With the same seed and the same prompt, the model makes exactly the same picture again. That's what lets me change one thing at a time and see its effect.

Level 0: six words

The prompt: a basketball player dunking in snow. The kind of line you type without thinking.

A basketball player in a navy jersey dunking in a snowy park, a hill and trees behind — Level 0, FLUX.2 [klein] 4B. Two arms, two legs, a hoop, a net. The letters on the jersey spell nothing.

No monster. FLUX.2 [klein] 4B, an open model with a permissive licence, made a believable dunk in the snow from six words. What I hadn't specified, it decided: a navy jersey, a hill, no particular framing. Its one visible miss is a classic: the letters on the jersey don't form a word.

Levels 1 to 3: saying what you actually mean

At each level I added one family of detail to the previous one. Same model, same seed, same size. Between 77 and 82 seconds a picture.

A basketball player in a red jersey leaping at a hoop in an indoor gym — Level 1, the player and the move: red jersey, both legs visible, ball above the rim.

A player in a red jersey dunking on an outdoor asphalt court dusted with snow — Level 2, the place: an outdoor court, fresh snow, flakes falling.

A player in a red jersey dunking on a snowy court, overcast light, background blurred — Level 3, camera and light: medium shot, 35 mm, f/2.2, overcast sky.

Level 1 surprised me. I described the player in detail, and the snow disappeared. The model put him in a brightly lit indoor hall. I hadn't said where, and by focusing on the player I'd left the setting blank. The cook filled the blank. That's the real lesson of this post: adding detail in one place can move something somewhere else.

Level 2 brought the snow back and put the court outdoors, because I wrote it.

Level 3 changed the framing and softened the background. The content stayed put. The jersey letters are still unreadable, and the backboard grew an odd second backboard underneath. As for the light, it was already grey at level 2, since the scene was in the snow. So I can't say my words about light changed anything.

The full prompts are under each picture, in "How this was made". Read in order, they follow the usual recipe: subject, then setting, then camera.

The negative prompt, and the model that ignores it

The usual recipe has a fourth step: the negative prompt, the list of what you don't want (deformed, extra limbs, bad anatomy, low quality). To see why it doesn't work everywhere, we need to talk about guidance.

Guidance is how strongly the model obeys the text. To apply it, many models compute two versions at every step, one that follows the prompt and one that follows the negative prompt, then steer away from the second. FLUX.2 [klein] runs at a guidance of 1: it skips that second pass. So the negative prompt is never read, and I can't show that step on this model. That's also why, with a model like this one, exclusions like "no text, no logos" have to be written into the positive prompt.

Stable Diffusion 1.5, older and smaller, uses a guidance of 7 and does read the negative prompt. And there, the six-word prompt does produce the promised defects. Left to right: level 0, level 3 without a negative, level 3 with one. 512×512, seed 2026, 63 to 68 seconds each.

A basketball player in white dunking in snow, one red shoe and one grey shoe, a lone pole beside him — SD 1.5, level 0: two different shoes, a pole holding nothing.

A player in a red jersey under a doubled hoop, an orange slab under the backboard — SD 1.5, level 3 without a negative: the hoop is doubled and melts into an orange slab.

A player in a red jersey reaching for a single hoop on a bent pole, mismatched shoes — SD 1.5, level 3 with a negative: one hoop, a bent pole, shoes still mismatched.

With the negative prompt there's only one hoop. It's tempting to call that proof it works. I can't. With one picture per prompt, the change could just as well be chance. Proving it would take twenty seeds and a count. I didn't do that.

A checklist to keep at hand

I use a five-point reminder when I write a prompt: subject, action, camera, light, and what to exclude. It's a habit of mine, nothing more. Here's what the pictures above really say about it:

To specify	Example	What these pictures show
Subject	Who, the outfit, the posture	Level 1 fixed the jersey and the pose, and let the model pick the place.
Action	What moves, what touches what	"Dunking" was enough from level 0. A still picture doesn't show motion.
Camera	Focal length, aperture, framing	Level 3 changed the framing and blurred the background.
Light	Source, materials	Can't be isolated: the light was already grey at level 2.
Exclusions	What you don't want	Ignored by FLUX.2 [klein]. Inconclusive on SD 1.5 with one seed.

What the model actually receives

Nobody should have to remember focal lengths to get a good picture. You can take what a person types and wrap it in a structured direction before it reaches the model. That's what we do in Orazaka, the open-source engine I'm building. The prompt under the picture below is exactly what the model received, in this order: the person's subject first, then the direction (light, lens, materials, palette), then the exclusions, written into the positive prompt because this model doesn't read a negative one.

A man in a white djellaba pouring mint tea from a silver teapot inside a Berber tent at golden hour — A short subject wrapped in a golden-hour direction. The full prompt, as the model received it, is under the picture.

The same prompt was used for a five-second clip, made by a hosted video model:

Hosted text to video: 134 s, $0.60. · 0.4 MB

How this was made

Model: Wan 2.2 T2V A14B · Apache-2.0
Engine: hosted: DigitalOcean serverless inference (steps and guidance not exposed)
Machine: hosted by the provider (not our machine)
Through: orazaka video --model wan2-2-t2v-a14b, routed by Orazaka
Prompt: Extreme close-up, slow motion: a stream of amber mint tea pours from a silver teapot held high into a small engraved glass, steam curling through a shaft of golden-hour light inside a Berber tent, dust motes floating, shallow depth of field, anamorphic lens flare, 35mm film grain, cinematic color grade.
Seed: 20261011
Size: 832×480 · 16 fps, 5.06 s
Time: 134 s, as recorded by the run
Cost: US$0.60 (the provider's price)
Orazaka run: 25948442-7c68-4e66-94d6-39cd8b96438d
Sound: music bed made by Orazaka (Studio video-sound, ACE-Step 1.5, MIT), -14 LUFS, AAC

What you can take away

A prompt doesn't need to be long. It needs to say what matters to you, because the model will decide everything else. When you add a detail, look at what moved elsewhere. And before you write a negative prompt, check that your model reads it.

What I didn't measure: variance (another seed could flip every conclusion), quality scored by anyone but me, and video, for lack of a plain-versus-detailed pair of clips. A hosted clip costs $0.60 and I didn't spend it on this. The times and prices of these models are in the post about costs.

Subscribe to the feed to follow new posts

The anatomy of an AI prompt: what each added line really changes

A prompt is an order you give a cook

Level 0: six words

Levels 1 to 3: saying what you actually mean

The negative prompt, and the model that ignores it

A checklist to keep at hand

What the model actually receives

What you can take away

Related posts

What it really costs to generate images and video: my bills, measured

An AI platform on one Mac: what breaks, what Spring AI 2 changed, and where the laptop gives up

Why I build sovereign AI in Montréal