Prompting line · stop 01 of 8 · 22 min · members
A structure that works across models: subject, treatment, optics, light, constraint.
Free with an account
Membership is free: an account opens all 86 script pages. The Lab, Studio Canvas and the paid guides need the $99 pass, paid once. Already signed in on this browser? The page opens by itself.
The problem
When everything is in one paragraph, you cannot tell which part is doing the work.
The typical prompt is a single run-on sentence containing a subject, four style references, three quality claims and a lighting note, in no particular order. It sometimes produces a good image. When it does not, there is no way to fix it except to rewrite the whole thing and hope.
That is the real cost of unstructured prompting: not that the results are worse, but that you never accumulate knowledge. Every prompt is a fresh guess.
Structure fixes this. If the prompt has named parts, a failure can be attributed to a part, and the fix is local. You start building actual understanding of what changes what.
The structure
Five blocks in a fixed order, each answering one question.
The order is not arbitrary. Models weight early tokens more heavily, so the subject goes first. Treatment establishes a category that the rest of the prompt modifies. Constraint goes last because it is a filter applied over everything above it.
Block one
A model renders materials and geometry. It cannot render an intention.
'A powerful image of confidence' gives the model nothing to draw. 'A woman in a matte black jacket standing three metres from the camera, shoulders square' gives it geometry, material and distance.
The test: could a set designer act on this sentence? If not, it is mood, not subject, and it belongs nowhere in the prompt.
Physical vocabulary that consistently works — matte, gloss, brushed, worn, woven, transparent, rough for surfaces; explicit distances and positions for arrangement; named colours rather than evocative ones. 'Bone white' beats 'ethereal pale'.
Scale is the most skipped detail
Models have weak priors about how large things are. Saying how far the camera is, or how big an object is relative to the frame, resolves a surprising number of compositions that were otherwise coming out wrong for no visible reason.
Blocks two and three
Two blocks that between them decide almost everything about how the image reads.
Treatment is one phrase naming what kind of picture this is: a technical documentation photograph, a fashion editorial frame, an architectural record, a product catalogue plate. This single phrase carries an enormous amount — lighting conventions, framing habits, colour handling — because these are real genres the model has seen thousands of examples of.
Choosing the treatment well does more than any amount of style-word stacking. One accurate genre name beats six artist references.
Optics then narrows it. A stated focal length works reliably; so does subject distance and an explicit depth-of-field instruction in plain language. Film stock names are strong handles on grain and colour response.
What does not work: aperture values, ISO, camera bodies, sensor formats. These are production metadata that rarely described the look in any caption the model learned from.
Block five
Three targeted exclusions beat twenty generic ones.
The final block names what must not happen — and specifically what this image is likely to get wrong, not a boilerplate list carried between projects.
If the frame contains a product, the constraint is that it keeps its shape and markings. If it contains type, that no text is invented. If it is documentation, that there are no logos or brands.
No logos, no readable text, no additional objects.
The subject stays off-centre with the lower third empty.Long negative lists dilute themselves. Every generic exclusion you add reduces the weight on the two that mattered, and several models handle negatives weakly enough that a twenty-item list is mostly decorative.
Using it
The structure only pays off if you exploit the fact that it is separable.
When an image fails, identify which block is responsible before touching anything. Wrong composition is subject or optics. Wrong feel is treatment. Flat and shapeless is light. Something appeared that should not have is constraint.
Change that block only, regenerate, compare. You are now running an experiment rather than making a wish.
Over time you accumulate blocks that reliably work — a lighting block that always produces shape, a treatment phrase that always lands the genre — and prompting becomes assembly from known parts. That library survives model changes far better than any complete prompt does, because the blocks are conceptual and the specific wording is not.
1:1 · two hours