Visual Skills bans the word cinematic and then counts nine fields where it says fourteen

Fourteen fields per storyboard row, nine of them written down

The seventh law is the spec sheet, and it has a mismatch in it. The text says fourteen fields per storyboard row and then enumerates framing, composition, camera, movement reason, eye-trace, duration, cut type, sound and light. That is nine. Five fields are unaccounted for, and the law's own rule about empty fields, that an empty field is missing direction, cannot be applied to fields nobody has named. The rest of the law is stricter than the count: per piece, exactly five anchors, one emotion, one motif, one object, one break, one final image. So a storyboard has to hit both a per-row requirement it cannot fully satisfy as written and a per-piece quota that is exact. An agent following only the enumerated list will produce nine fields per row and still pass whatever check is looking for the five anchors.

Murch's six cuts arrive with percentages attached

The editing law reproduces Walter Murch's Rule of Six as a priority order where each item outweighs everything below it combined, and attaches a weight to each: emotion at 51%, story at 23%, rhythm at 10%, eye-trace at 7%, screen plane at 5% and 3D space at 4%. The document then makes one point about the ordering, that cutting for pace is item three, and that serving item three ahead of emotion and story is how the mush it names gets made. The claim about models is direct: this reordering is the default behaviour of every model you will ever prompt. The six numbers are presented as a bar chart in a preformatted block, with no citation attached to where they come from, and they sum to exactly 100, which reads as a weighting rather than as measured data from any edit. Treat them as the skill's priorities, not as a measurement.

The Details Law bans the adjectives the models reward

The second law requires every shot to own three physical facts: one environmental pressure such as cold refrigerator light or wet asphalt, one micro-action of the body such as a locking jaw or whitening knuckles, and one sound anchor or visual motif. The example given for why this matters is short: he is sad does not render, a jaw does. The banned word list that follows is where the skill is bluntest. `cinematic`, `epic`, `stunning`, `masterpiece`, `beautiful lighting`, `dynamic camera` and he is sad are all banned everywhere in the skill, on the grounds that each is a placeholder for a detail the writer failed to invent and that not one of them renders. That is a testable claim about what the models do with adjectives, and it is the load-bearing claim of the whole document, since the syntax tables underneath only matter once the prompt says something.

The bad prompt column is the more useful half of the example

The document puts two prompts side by side. The one everyone writes is four lines and no facts:

cinematic shot of a man
in a kitchen at night,
epic lighting, moody
atmosphere, 4k

The annotation under it names what is missing: no desire, no obstacle, no geometry, no cut, no final image, and a note that the model picks all five itself and picks differently on every run. That last clause is the argument for the entire skill, since it turns re-running the prompt into a lottery rather than an iteration. The skill's version of the same scene opens with the anchors:

Emotion: hunger as loneliness. Object: last sausage.
Final image: fridge light dying on his face.

Shot 1 · 0.0-1.6s · wide, 24mm, static
Dark kitchen. He stands with one hand on the fridge
door, not opening it. Only the wall clock moves.

and the annotation claims every line is a physical fact a camera could record, with a reason for the move, a body carrying the emotion, a sound, an object and an ending. Read together with the law that bans cinematic and epic, the two columns are the same argument: the adjectives are what the writer writes when nothing has been decided.

The output is gated twice and a failure is not returned

The rules are described as non-optional. The dramaturgy reference loads before any model file, so the rules are in context ahead of the syntax rather than after it. Then the output is gated twice on the way out, by a six-point dramaturgy check and by a three-detail audit on every shot, and a prompt that fails either is not returned. The checklist line itself is cut off in the copy that reaches this page: it names scene formula, three details, three jobs, motivated camera, and then breaks off partway into a fifth entry beginning with r, so the six points cannot be counted from what is visible. The first law is the one that anchors the rest, the scene formula of desire plus obstacle plus geometry plus gaze plus rhythm, each named in a single sentence before a word of prompt is written, with the claim that anything less is decoration.

Rhythm is a staircase, and the pause is the load-bearing step

The sixth law treats montage as a staircase: long, then shorter, then shorter, then a pause, then the impact, with the claim that the pause before the hit matters more than the speed of the cuts. It supplies beat maps for 15, 30, 60 and 90 second clips, named Hook, Pressure, Crack, Impact and Aftermath, with one instruction attached: never skip the Crack. The fourth law is the deletion rule that pairs with it. A shot has to change emotion, advance action, or increase pressure, and a shot that does none of those is deleted however pretty it came out, with a beautiful establishing shot named as the canonical example of a shot with no job. The fifth law adds staging by director: a Fincher camera move must answer what changed or stay static, a Spielberg frame must keep the hero, the threat and the exit findable even in chaos, and a Kurosawa scene runs on one weather and one pressure.

Version-specific syntax with no releases to signal a version bump

The repository is two skill directories, `video/` and `image/`, plus a `.claude-plugin/` manifest, a LICENSE, a NOTICE, a Russian README alongside the English one, and an assets folder. The project description claims exact prompt syntax for Seedance 2.5, Kling 3.0 Turbo and Omni, Veo 3.1, Nano Banana 2 and GPT Image 2.5, so the syntax half of the skill is keyed to vendor version numbers rather than to stable interfaces. There are no GitHub releases, no changelog and no version file in the tree, so a vendor shipping the next model version produces no signal here beyond a commit, and the last push is dated 2026-09-16. The install path is the Agent Skills standard rather than a package manager, which suits prose instructions that carry no code to break, but it also means nothing tells an agent which version of the rules it has loaded.

Editorial conclusion

Read this as prompt engineering with a checklist attached, because that is what it is: a CC-BY-4.0 rulebook that an agent is told to load before any model file and that refuses to return a prompt failing its own audits. The rules are worth stealing regardless of which model you use, since the Details Law and the three-jobs rule are model independent. Two caveats. The syntax tables are keyed to model version numbers such as Kling 3.0 and Veo 3.1, and there is no release or changelog to tell you when a vendor moves past them, so verify the syntax against the vendor's own current documentation. And the seventh law promises fourteen shot-card fields while listing nine, so an agent working from that section alone will under-specify the storyboard.

Frequently asked questions

Official sources

1. Issues
2. License: CC-BY-4.0
3. Project website
4. README
5. smixs/visual-skills on GitHub