The first poster came back with lettering on it that no alphabet contains. Eilidh Ramsay had asked for the name of a fiddle festival across the top of a harbour scene, and what arrived was a confident row of shapes with the weight and rhythm of words, spelling nothing.
That is the species of failure to expect on this side of the subject. Nothing was wrong in the sense of being false. The picture was competent, the light was good, and the part that failed failed by looking exactly like the thing it was not.
The request that made the poster
Here is what Eilidh asks for now, whole, with the frame set before anything else.
- Two to three portrait, and compose for that ratio rather than cropping to it.
- A small stone harbour at dusk, three-quarter view, seen from the water.
- Two fishing boats tied up along the quay, the nearer one with a red hull.
- Screen printed poster: four flat colours, visible paper grain, no gradients and no glow.
- Low sun from the left, long shadows crossing the quay.
- Leave the upper third of the frame empty.
- No text, no lettering, no signage anywhere in the image.
The ratio is decided first because it is the one thing a reframe cannot add: a square image cropped tall loses the quay, and no amount of editing puts back what was never drawn. The red hull is doing quieter work. It is a particular, and it is what makes two runs of this prompt comparable to each other, because a picture described only in general terms varies in a hundred places at once.
The fourth line is the one that changes the most, and it is written in printing words rather than art words. Left out, the default arrives: something photographic, softly lit, with a gradient in the sky, because that is what the greatest number of images look like. Four flat colours and no gradients is a constraint a printer would recognise, and it is more useful than any adjective about mood.
The last two lines exist because of the lettering. The model draws letter shapes, not letters, so set the type as a separate layer. Ask for an empty upper third and the poster arrives with somewhere to put a name that a person will set, in a typeface somebody chose, at a size that will still read on a village noticeboard.
Three things that only mean anything together
Eilidh liked the fourth image of an evening and could not find it again the next morning. Keep the seed, or the version you liked cannot be found again: the seed is the starting noise the picture was refined out of, and without it the same words produce a different harbour every time.
The seed alone is not enough either. The same prompt and the same seed on a different model version is a different picture, so the three have to be recorded as a set. Once they are, the actual working method becomes available: change one term at a time with the seed held fixed, or nothing is comparable. Swap dusk for overcast noon with everything else pinned, and the difference on the screen is what that word does, which is the only way anybody ever learns what the words do.
A style with somebody in it
The temptation on the first evening is to end the prompt with the name of a painter whose posters Eilidh admires, who is alive, and who sells prints. Name a living artist and you have named a person, not a style setting. The honest version of what she wanted was in the medium clause all along, and it took three sentences of screen printing vocabulary rather than one surname.
The same line runs through the other outputs and gets sharper as it goes. Speech and sound from a description is where it is hardest to hedge, because a voice that is nearly somebody's belongs to somebody, and the festival's radio spot was read by a volunteer for that reason. On the moving side, generated footage of a real place is a drawing of it, and a caption should say so: the harbour in the video is not the harbour, and an audience that has been there will know before the caption does.
The order
Learn the steering where an attempt costs four seconds
Everything transferable about steering an output is cheapest to learn on still images, and then spends well everywhere else.
- start hereImagesA change and its effect, side by side, in seconds01
- 02WritingThe one most people need, and a rewriting problem more than a prompting one
- VideoWhere generated help with real footage beats generated footage03
- 04Voice and audioWhere the person behind the output stops being an abstraction
- CodeThe only output a machine can check, and trusted for that reason too readily05
There are no dates on this and no weeks implied. It is an order and nothing more; two of these may never be reached at all.
The order to take these in
Start with prompting, steering and editing a picture, whatever you actually want to make, because it is the one where a change and its effect sit side by side in four seconds. Everything transferable about steering is learned there cheaply and then applied to media where each attempt is slow or expensive.
Drafting without sounding like everyone else comes next and is the one most readers need most, and it is not a prompting problem so much as a rewriting one. Generated footage, and generated help with real footage is third, where the second half of that name is the part with the working uses in it. Generated code, and reviewing it properly sits apart from all of them: it is the one output that a machine can check, which makes it the easiest to trust for the wrong reason.
The stall is the reroll
People stop making progress here at the moment rerolling becomes cheaper than thinking. The image is not right, the button is there, and pressing it produces a different image in four seconds, which feels like iteration and is closer to a slot machine. It does not feel like a stall because output keeps arriving.
The tell is being unable to say what any single word in your own prompt is doing. After a thousand rerolls a person has a thousand images and no method; after twenty deliberate single-term changes on one fixed seed they have a working vocabulary they can use on the next brief. Eilidh's harbour took eleven attempts, nine of which changed exactly one clause.
What to open in 2026
ComfyUI is where the seed, the sampler and the model version are visible rather than hidden, free and open source as of 2026, and it replaced a text box that throws the settings away when you close it. It writes the whole recipe into the image file, which is the reproducibility problem solved by storage instead of by discipline.
Content Credentials, the C2PA standard, attach a signed record of how a file was made and travel with it as of 2026, replacing a caption that nobody could check. The specification is open and free, support in cameras and editors is growing rather than universal, and the honest position is that it proves what was declared rather than what is true.
Inkscape sets the type, free and open source as of 2026, and it replaced asking the model for the lettering and receiving shapes that are nearly letters. The poster is a generated image placed under vector type, which is also how it stays editable when the lineup changes in March.
All of which leaves the question the empty upper third only postpones. Eilidh can tell a good poster from a bad one because she has seen a thousand posters, and the steering above is worth nothing to somebody who cannot. The judgement is the scarce part and it did not come from any of these tools, which is the argument for marks, layouts and printed things sitting next door. The narrower version of the same question, which is how to tell a good output from a plausible one before acting on it, is getting a useful answer instead of a plausible one.