Charge every word a line.
Below is a real prompt of the kind an LLM writes when you ask it for one: thorough, confident, and mostly furniture. It was written against these four requirements — a cutaway drawing of a lock, for a page of documentation. Click any word to see what it is actually doing. Eight of the thirty-two are working.
Reads as a pin tumbler lock, opened along its length.
The pins and springs read as the mechanism rather than decoration.
Ink on paper, matching the other plates in the document.
The labels are added afterwards in HTML, so the drawing arrives bare.
Every word gets one line: what it does, and which requirement it serves. A word nobody can write a line for is a word to cut — which is the whole mechanism, because it makes padding cost something.
19 words cut
Intensifiers, grammar, and instructions telling the model how to do its job. None of them named anything that had to appear.
One negation replaced
no text or labels anywhere
bare margins
A negation names the thing it forbids, and the model draws what it hears. A described bare surface excludes the lettering by itself.
One requirement added
five pins and springs seated over the plug
It was the second requirement all along, and the draft never said it. Padding hides that; a ledger cannot.
The trim is measured against the requirements, never against the word count. A shorter prompt that drops a requirement is the worse prompt — and the requirement that goes first is always the one that was hard to render.
Length is not the dial.
The requirements that were hard to satisfy fall out quietly, because they were the hard ones. Push further and the words stop holding together at all.
Now the prompt fills with the writer's own map of meanings: the justifications, the hedges, the explanation of the word it wrote a moment ago. All of it gets rendered.
The reasoning is real work and it has to go somewhere. Given a room of its own it stops competing with the words that decide, and the prompt can finally be measured.
So the skill changes the shape of the answer rather than handing out rules about wording. Four sections, and the image model only ever sees the second one.
Four sections, and one of them is the mechanism.
Ask for a generation prompt and this is what comes back. Section 3 is where the discipline lives: it is the bill for section 2.
Requirements
Bullets. What the picture has to achieve, written before the prompt exists. This is what a trim is measured against, so the hard-to-render requirement stops being the one that quietly vanishes.
Read by you
The prompt
The literal text, ready to paste. Everything in it is there to be rendered — which is a strong claim, and section 3 is where it gets tested.
model ever sees
Every word, justified
One line per word, in order: what it does, what it connects to, which requirement it serves. Words, not phrases. Paying a line per word makes padding cost something, so it stops being free — and the map of meanings finally has a room that is not the prompt.
for section 2
Notes
Free space. What was tried, what the image model keeps getting wrong, alternatives considered, what is still open. A result that comes back wrong lands here, and the fix is found upstairs in section 3.
worth keeping
Seven things that move a result.
Observations rather than commandments. Each has exceptions, and the skill leaves those to your judgement.
A negation draws the thing it forbids.
The words are all the model has, and yours just said lettering. Describe the condition you want and it excludes by itself.
The pile-on is the expensive habit.
A wrong result invites a longer list of what must be absent. It arrives disguised as diligence and it makes things worse. The fix is a better positive sentence.
Name what everyone pictures alike.
A shared name fixes a design in one word. A precise but obscure term resolves to something plausible and wrong, and four careful sentences each fetch the nearest common object instead — so describe the shape of those.
Saying it twice is weight.
Repetition is counted rather than ignored, so the second copy competes with the words that decide. One statement, in the strongest words you have.
Some words carry a prior you lose to.
Where the model holds a strong expectation about what belongs on a surface or in a role, arguing feeds it. Describe the alternative and let it stand on its own.
Some words name an artefact, not a subject.
Character sheet fetched a games-industry asset — flat vector fills, nothing drawn — and no amount of describing the character reached past it. Two words came out and the picture arrived.
The structure is the whole of the discipline.
Rules about wording go stale the moment an image model is retrained. The four sections do not, because they change where the reasoning goes rather than what it says. Everything the structure does not decide stays yours.
However long the picture needs. Fewer words have won consistently, and a requirement is worth more than a short prompt.
A list or a table, whichever reads better. Small connecting words get as much room as they need and no more.
The picture is the test, and you are the one looking at it. What the attempt taught goes into the notes and the next draft starts from there.