The Anatomy Of A Prompt That Still Works Next Quarter
Subject, action, setting, camera, light, style. Six slots. Fill them in that order and most of what people call prompt engineering disappears.
Most bad prompts are not badly written. They're incomplete in a specific, predictable way: they describe the subject and nothing else, and then the model invents the other five things.
When the same prompt gives you a different result each time, that's usually not randomness. It's the model filling in slots you left empty.
Six slots
Order them roughly like this. Not because the model parses positionally, but because writing in this order stops you skipping things.
1. Subject — who or what, and enough detail to be identifiable.
a welder in a heavy canvas jacket
2. Action — what they are doing. One verb. Present tense.
lowering his mask
3. Setting — where, and what is around them.
in a cluttered workshop, sparks on the floor, tools on a pegboard wall
4. Camera — size, angle, lens, and for video, movement.
medium shot, slightly low angle, 35mm, camera static
5. Light — direction, quality, colour.
harsh blue arc light from the torch, deep shadow elsewhere
6. Style / medium — the register of the image itself.
shot on 35mm film, visible grain, muted colour
Assembled:
A welder in a heavy canvas jacket lowers his mask in a cluttered workshop, sparks on the floor, tools on a pegboard wall. Medium shot, slightly low angle, 35mm, camera static. Harsh blue arc light from the torch, deep shadow elsewhere. Shot on 35mm film, visible grain, muted colour.
Nothing in there is clever. Every clause is a decision that would otherwise have been made for you.
Describe the observable, not the intention
This is the one principle underneath everything else.
Models render what can be seen. "Powerful", "premium", "engaging", "eye-catching" cannot be seen — they're conclusions a viewer draws from things that can. Write the causes, not the effect.
| Intention | Observable |
|---|---|
| Make it look premium | Matte black surface, single hard key light, deep shadow, no clutter |
| Make him look powerful | Low angle, camera below eye level, subject fills upper frame |
| Make it feel cosy | Warm practical lamps, shallow depth of field, soft falloff |
| Make it eye-catching | High contrast, one saturated colour against neutral surroundings |
The left column produces a different result every run. The right column produces the same one.
Say what you want, not what you don't
"No text, no watermark, not blurry" is a habit carried over from older image models with dedicated negative prompt fields. In a single prompt string, naming a thing tends to raise its salience rather than suppress it.
If a model supports a separate negative prompt field, use that field. If it doesn't — and most of the video models in common use don't expose one — convert the negative into a positive:
- Instead of
no clutter→bare surfaces, empty background - Instead of
not blurry→sharp focus throughout, deep depth of field - Instead of
no people→empty street at dawn
Specific beats long
There's a persistent belief that longer prompts are better prompts. They aren't. Length only helps while every added clause is a decision. Past that point you're adding noise, and on several models you're also risking silent truncation — the prompt caps vary enormously and models generally don't tell you when they've cut you off.
A useful test: delete any clause and ask whether the output could now come back meaningfully different. If not, the clause was decoration.
Padding that does nothing:
a beautiful, stunning, highly detailed, masterpiece, 8k, ultra realistic, award-winning photograph of a coffee cup
Every one of those adjectives is a quality claim rather than a description. The model has no coherent visual definition for "masterpiece". Replace all of it:
A white ceramic coffee cup on a walnut table, morning light from a window on the left, shallow depth of field, steam visible against the dark background.
Shorter, and it actually specifies an image.
What changes between models, and what doesn't
The six slots survive model upgrades. What doesn't survive is anything you learned about a particular model's quirks — the phrasing that happened to work around a weakness in one version usually stops mattering in the next, and occasionally starts hurting.
This is the practical reason to keep the two separate in your own notes. Structure is durable knowledge. "Seedance responds better if you put the camera clause first" is a workaround with an expiry date, and it needs re-testing every time the version number moves.
A checklist
Before you send a prompt, check you've said something about each:
- Subject — identifiable, not generic
- Action — one clear verb
- Setting — where, and what's around
- Camera — shot size at minimum
- Light — direction and quality
- Style — medium, film stock, or render register
Any box you leave unticked, the model ticks for you.
Keep reading
Change One Thing: How To Iterate On A Prompt
Rewriting the whole prompt after a bad result teaches you nothing. Changing one clause at a time is slower per attempt and far faster overall.
Your Prompt Is Probably Being Truncated And Nothing Told You
Prompt caps across current image and video models range from 1,000 to 20,000 characters. Go over, and the tail is usually dropped in silence.
Why You Should Generate The Image First, Then Animate It
Text-to-video gives you one roll of the dice on everything at once. Image-first splits it into two cheap decisions you can actually control.
