Structure beats phrasing
Most prompt advice is folklore about magic words. Lists of adjectives, a camera body nobody used, a suffix somebody found once and now copies everywhere. Some of it works on one model this month. Almost none of it survives a model update or a colleague.
What does hold up across models and across revisions is structure: subject, then framing, then light, then treatment — each stated once, in the same order every time.
Written that way a prompt becomes editable. Somebody can change the light without accidentally rewriting the subject, because the light is its own clause rather than an adjective buried in a sentence about a person. That is the entire difference between a prompt that is a record and a prompt that is a lucky incantation.
- Subject
- What is in frame, in the fewest words that identify it. Detail that belongs to a reference image should be in the reference, not restated here — restating it is how the two start to disagree.
- Framing
- Shot size and angle. This is the clause you change most often, which is exactly why it stays in its own place instead of being tangled into the subject.
- Light
- Direction, quality, time of day. Shared across a scene rather than restated per shot — and in a graph, that means it belongs upstream.
- Treatment
- Medium, grade, lens character. The part that makes a set feel like one set rather than several similar images.
Where prompts live in a graph
In the canvas a prompt belongs to a node, alongside the settings and the result it produced. That is what makes it a record rather than a message — you can look at a frame and see exactly what asked for it, months later.
Shared clauses belong upstream. If four shots share a light, that light is a node they draw from, not a sentence copied four times that will drift on the second edit. This is the single most useful habit to build, and it is the one that separates people who find the tool powerful from people who find it fiddly.
Habits worth building
Write it once, then cut it
First drafts are always too long. Remove everything the reference image already says. What is left is usually half the length and produces better results.
Change one clause at a time
Two changes and a better result teaches you nothing about which one helped. This is slower for one image and much faster across a project.
Keep the failures
A node's run history includes the versions that did not work. Those are more informative than the one that did, and deleting them removes the evidence.
Name things the way you speak
You will read this again under deadline. A node called "warm afternoon key" is findable; one called "test 3 final" is not.
Questions
Do you publish a list of prompts I can copy?
Deliberately not. A list of phrasings is exactly the folklore this page is arguing against, and it goes stale with every model release. The templates and the workflow library show prompts in the structure that makes them reusable, which is the transferable part.
Do negative prompts help?
Model-dependent, and less than people think on current models. Try removing yours before adding to them — a long negative prompt is often compensating for a subject clause that says too much.
How long should a prompt be?
Shorter than your first draft. Beyond a certain length most models weight the beginning heavily and effectively ignore the end, so the last third is frequently doing nothing at all.
Does the same prompt work across models?
The structure does. The specific weighting does not, and that is fine — a graph lets you change the node rather than rewrite the project.