Guide
Why the model ignores half your prompt
You asked for six things and got four. Adding the missing two in capitals, or at the end, or three more times, mostly does not work — and the reason it does not is worth understanding before you spend an evening on it.
The mechanism
Attention is finite and spread unevenly
A prompt is not a checklist the model works through. It is turned into a single bundle of meaning, and every part of the image is generated with reference to that whole bundle at once. Nothing tracks which of your requirements have been satisfied, because there is no list of requirements — there is one blended instruction.
That blending has limited capacity. Add more concepts and each one gets a smaller share of it. At four or five ideas most of them come through; at fifteen, several will be approximated and one or two will vanish entirely, and which ones is not under your control in the way you would expect.
So the failure is not the model refusing an instruction. It is the instruction arriving faintly, mixed with everything else you said.
The practical rule
Position beats emphasis
Words early in a prompt generally carry more weight than words late in it. This is the single most useful thing to know, because it is the opposite of how people edit: when something is missing, the instinct is to append it, which puts it in the weakest position available.
Move it to the front instead. If the cracked window is the point of the image, the prompt starts with the cracked window, not with the lighting and the lens and the mood. The things you care most about go first and the atmosphere goes last, where losing some of it costs you nothing.
Capitals do nothing. Exclamation marks do nothing. Repeating a word three times does something, but it is blunt and it steals capacity from everything else — you will often get the repeated thing and lose two others you had already.
The second rule
Shorter prompts obey better
A prompt of eight well-chosen words is usually followed more faithfully than one of eighty. This runs against the habit of pasting in long stacks of quality tags, and it is worth testing yourself rather than taking on trust: take a prompt that is not working, delete two thirds of it, and see what comes back.
Most long prompts contain a lot of nothing. “Masterpiece, best quality, highly detailed, 8k, award winning” is five concepts competing with your actual subject for the same finite attention, and contributing very little in exchange. Modern models do not need to be told to try hard.
Where length genuinely helps is specificity about one thing — a precise description of a single subject beats a vague one. Where it hurts is breadth: many requirements, each lightly specified.
Why it hits identity work hardest
Your character is competing with your scene
This is the quiet reason character consistency gets worse as scenes get more elaborate. A face described in words is already a weak instruction; put it in a prompt that also specifies a location, an outfit, a camera angle, a time of day and a mood, and it is one voice in a crowd.
It also explains a pattern people find baffling: the character was holding fine, then they wrote a richer scene and the face went. Not because the scene description was wrong, but because it was expensive. Every word spent on the room was taken from the person in it.
Which is the argument for carrying identity outside the prompt entirely. Conditioning on a reference image does not compete for prompt attention at all, so the whole prompt is free to describe the scene — which is both why it holds better and why it leaves you more room to write.
What to do instead
Split the work across passes
When a prompt has too much in it, the answer is usually not a better prompt. It is two passes. Generate the composition you want, then fix the specific thing that came out wrong by editing that region — a second generation that only has to care about one requirement will satisfy it.
Negative prompts are worth a mention because they are widely misunderstood. They are useful for pushing away a recurring unwanted quality, and they are not a way to add a requirement. “No hat” is a reasonable negative; trying to get a specific pose by listing every pose you do not want is not going to work, and the list costs capacity.
Last: change one thing at a time. Prompting feels like it should reward big edits, and it punishes them — when you change five words and the result changes, you have learned nothing about which word did it.
Keep reading
The rest of the series
- Image-to-video or text-to-video?The choice decides whether you control the first frame or the model does.
- Why AI video looks floatyThe weightless, slow-motion tell: where it comes from and which parts you can actually prompt away.
- Why hands and faces come out wrongTwo different failures with two different causes - and why asking for perfect hands does nothing.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.