The output is wrong and the instinct is to rewrite the whole prompt. That is the slowest way to fix anything. Read the failure first — it usually tells you which line to touch.
- Change one thing at a time. Rewriting everything means you learn nothing and cannot repeat the fix.
- Most failures are one of six kinds, and each points at a different line in the prompt.
- Ignored instruction usually means competition — something else in the prompt contradicts it, or buries it.
- Wrong subject is usually word order. Earlier words carry more weight in most image models.
- If it changes wildly every run, you are underspecified — the model is filling a gap you left.
- If it will not change at all, something is anchoring it — a seed, a reference image, or a term with a very strong pull.
A bad generation is not random. It is the model resolving a conflict or filling a blank, and it does that the same way every time.
So the question is never "how do I write a better prompt". It is "which of those two happened, and where".
The method, before the symptoms
Three rules that make everything below work. Skip them and you will fix things by accident and be unable to do it again.
Diagnosis is an experiment, and an experiment with two variables is not an experiment. The discipline is the technique.
Six failures and what each one means
It ignored an instruction completely
Most likely cause: something contradicts it. "Empty street" and "bustling market" cannot both be satisfied, so one wins. Read your prompt back looking only for pairs that fight.
Second cause: it is buried. An instruction in the middle of a long paragraph carries less weight than one near the front. Move it, do not repeat it.
Third cause: the model cannot do it. Exact text, precise counts of objects, and specific hand positions fail across many models regardless of phrasing. If three rewrites change nothing, stop rewriting — you have found a ceiling, not a wording problem.
It rendered the wrong thing as the subject
Cause: word order and weight. Many image models weight earlier tokens more heavily, so whatever you mentioned first tends to become the subject. If the background became the star, the background was described first or described in more detail.
The fix is proportional, not additive: cut the description of the thing that took over. Adding more words about your intended subject makes the prompt longer and the competition tighter.
Every run looks completely different
Cause: you left a gap and the model filled it. Wild variation between runs is a map of what you did not specify — if the location changes every time, you never fixed the location.
Diagnose it deliberately: generate four and list what stayed the same. What stayed is what your prompt controls. Everything else is the model's choice, and it will keep choosing.
Nothing you change makes any difference
The opposite failure, and more frustrating. Something is anchoring the output.
Usual suspects, in order: a fixed seed you forgot you set; a reference image overriding your text; a very strong style term dragging everything with it; or an edit applied to a cached result rather than a fresh generation.
Test it with a deliberately absurd change. Ask for a completely different subject. If the output barely moves, the anchor is real and it is not your wording.
It looks right but feels wrong
Nothing identifiable is broken and the image still reads as fake. This is almost always lighting or surface, not composition.
Check the shadows agree on one source, and check whether every surface is unnaturally clean. Both have their own page: why AI output looks fake covers the mechanism and the specific words that undo it.
Video: it drifts, morphs, or slides
Drift and morphing mean too much is changing per second. The model is inventing more than it can keep consistent. Reduce the movement before you touch anything else — one significant movement per shot.
Feet sliding, objects without weight, hair moving independently are physics failures, and they respond to describing the physical event — what touches what, what takes weight — rather than to more style words.
Identity slipping between shots is an anchoring problem, and it is what identity-lock instructions and reference binding exist for. Character consistency covers the methods.
The order to change things
When several things are wrong at once, fixing them in this order stops your fixes fighting each other.
Most people work this list backwards — they adjust style first, which changes composition and light as a side effect, and then cannot work out why the fix stopped working.
Three rewrites with no movement means it is not a prompt problem. At that point the options are a different model, a reference image, or fixing it after generation in an editor.
Prompting is not the only tool and it is not always the cheapest one. A thirty-second edit can settle something that twenty generations will not.
This is craft reasoning, not measurement. Nothing here is a controlled comparison — it is a description of how these systems behave and what that implies about where to look. The one claim that rests on vendor documentation is token weighting, which Black Forest Labs states for FLUX: earliest words are weighted most.Black Forest Labs, FLUX.2 prompting guide, docs.bfl.ai. Weighting behaviour differs between architectures and is not universal across models. Checked 31 Aug 2026
Models differ, versions change, and a technique that works on one may do nothing on another. The method — one change at a time, fixed seed, subtract first — is what transfers. The specific fixes are starting points to test on yours.
The through-line: the output is telling you which line is wrong. A failure that changes every run is a gap; a failure that never changes is an anchor; a failure that ignores you is a contradiction.