The Prompt Skeleton

A prompt is not a sentence

The single biggest gap between someone getting lucky with a video model and someone directing it reliably is prompt structure. A one-line description — "a person walks into a room and looks surprised" — leaves the model to fill in everything you didn't specify: what room, what walk, what kind of surprised, how long, what happens after. Every gap becomes an opportunity for the model to improvise something you didn't want.

The fix is a structured timeline prompt: four distinct blocks, always in the same order, each doing a job the others don't. It isn't decoration — each block exists because it fences off a specific way generations go wrong when that information is missing.

CONTEXT: the world, stated once

CONTEXT sets the scene before any motion is described: who the subject is, what they're wearing, where they are, what mood the shot should carry, what the color logic looks like, and the overall visual style. This block gets written once and holds for the entire shot — you're not re-describing the subject at every beat, you're establishing the world the beats happen inside.

Color logic deserves special attention here. A vague "cinematic lighting" note does less work than a specific description of what's warm, what's cool, and where the eye should land first. The more precisely CONTEXT nails the visual world, the less room the model has to substitute its own defaults.

SHOTLIST: directing time

SHOTLIST is a timestamped beat list — a short, ordered sequence of what happens and when, written as ranges like 0-3s, 3-7s, 7-11s. This is where you direct time inside the world CONTEXT established.

Two disciplines matter here that are easy to violate by accident. First: don't over-cram. Packing nine beats into a fifteen-second shot doesn't produce nine faithfully compressed beats — it produces a generation that drops several of them, and you don't get to pick which ones survive. Fewer, clearer beats consistently outperform a packed timeline. If a scene genuinely needs more coverage, that's a signal to split it into more than one generation, not to compress harder.

Second: include a "one continuous shot, follow logic" instruction. Left unconstrained, a video model will sometimes invent its own cuts mid-generation — a hard edit you never asked for, breaking the continuous take you intended. An explicit instruction that the shot is one continuous take, following the logical progression of the beats you wrote, closes that gap.

RULES + NEGATIVE: fencing off bad habits

This block is where you state hard constraints and negative prompts explicitly, rather than hoping the model infers them from context. Two rules matter enough to be non-negotiable in this discipline:

Three characters, maximum, per generation. Above roughly three named characters in a single shot, failure modes become common and repeatable: characters randomly falling, taking a wrong turn mid-motion, general glitchiness across the frame. This isn't occasional bad luck — it's a demonstrated ceiling. If a scene needs a crowd or an ensemble beat, split it across generations or composite the result, but don't fight the ceiling by naming a fifth character and hoping.

Never depend on in-scene text. Signs, dashboards, menus, numeric readouts — any text the model has to render inside the generated frame is unreliable, even when your start frame shows it perfectly legible. The professional move is to never route legible information through the model's own text rendering at all. Generate the shot, then add the text as a post-production overlay compositing pass. This single habit eliminates an entire category of reroll.

A worked RULES + NEGATIVE block looks like this:

RULES + NEGATIVE:
- One continuous shot, follow the shotlist's logical progression — no invented cuts
- Maximum 3 named characters in this generation
- No on-screen text, numbers, or legible signage — composited separately in post
- Negative: no morphing, no extra limbs, no logo drift, no wardrobe color shift mid-shot

REFERENCES: tagging what each image means

The REFERENCES block tags every reference image by the role it plays — "image one = subject," "image two = wardrobe," "image three = prop." Without this tagging, a model receiving multiple reference images has to guess which one governs which aspect of the shot, and it guesses wrong often enough to matter. Tagging removes the guess.

Generate-many-pick-best is the job, not a failure

It's tempting to treat rerolling as evidence a prompt failed. It isn't. Generate-many-pick-best is the standard professional workflow for this discipline — budget rerolls into every scene from the start, the same way a live-action shoot budgets multiple takes. A well-structured prompt with a locked reference will still produce variation across attempts; the job is picking the best one, not demanding the first one be perfect.

Direct the emotion, don't assume it

Emotion adherence is a genuine strength of current video models — but only when you direct it explicitly. A shot description that only covers physical blocking ("walks toward the window, stops, turns") gives the model nothing to say about how the character should feel doing it. Add the emotional arc directly into the SHOTLIST or CONTEXT: "starts guarded, shifts to open relief across the push-in." Models that are given an explicit emotional throughline tend to honor it convincingly — leaving it implicit is the mistake, not the model's limitation.

One closing note on voice: for hero content — the shots that carry your subject's actual dialogue or narration — a real recorded voice beats an AI voice clone whenever it's practical to record one. Voice cloning has real uses, but nothing currently beats an actual human take for the content that's meant to represent that human directly.

Build it: assemble the structured prompt

You're going to implement the function that turns a scene spec into a well-formed structured prompt string — CONTEXT, then SHOTLIST, then RULES + NEGATIVE, then REFERENCES, in that order, with a guard that warns when a scene names more than three subjects.