Every official prompting guide across the major video models converges on the same handful of things a good shot prompt needs — and disagrees on the order and the syntax. This is the shared internal shape, nine ready-to-fill templates built from it, and the specific syntax differences that trip people up when they carry one model's habits to another.
The universal slot skeleton
Eight content slots render into the prompt; two more are call-shape settings that mostly live outside the prose entirely:
The 10-slot skeleton
1 Framing (shot size + angle) — 2 Subject (identity, pasted verbatim from a locked reference) — 3 Action (one verb phrase, one intention) — 4 Setting (place + time of day) — 5 Camera (one move, one speed, or an explicit lock) — 6 Light (source + quality + direction) — 7 Look (one style anchor + grade) — 8 Audio (ambience/SFX/dialogue — never leave this empty) — 9 Anchors (which reference images, in which role) — 10 Constraints (positive states; a ban-list only if the model has a real negative-prompt field).
Nine templates for every shot type
1 — SPEAK: one person, one line
Framing + Subject (identity card, verbatim) + Action (verb, weight on a named foot, eyes to a named target, one involuntary beat) + Setting + Camera + Light + Look + Audio (one line, one speaker only, delivery style, "mouth closed after the line," the ambience bed, "no other speech") + Anchors + what everyone else in frame does while listening.
e.g. Keep the line to one speaker and roughly 3-5 seconds of speech — longer lines desync, and two speakers in one generation causes line bleed.
2 — REACTION: the cheapest, most reliable shot you have
Close-up, shallow depth of field + Subject + Action (an attention verb toward a named off-frame thing, then one micro-move) + Setting + Camera (locked or a very slow push) + Light + Look + Audio (ambience only — no dialogue from anyone in frame) + Anchors.
e.g. The defining move is naming the off-frame object of attention — the thing being reacted to stays out of shot, so the face is the whole story.
3 — INSERT: an object is the subject
Macro or extreme close-up + Subject (the object, its material and state) + Action (what changes about it) + Setting (the surface it sits on) + Camera (one slow push, or locked) + Light + Audio (the sound the object makes, plus the bed) + Anchors (a prop reference).
e.g. Render this one terse — fragments, no connective words. Inserts are where brevity wins outright.
4 — ESTABLISH: the place is the subject
Extreme wide or high angle + Subject (optional, small in frame if present) + Setting + Camera (a reveal move) + Light + Look + Audio (the bed this place makes) + Anchors (a location reference).
5 — SILENT ACTION: the only template that gets multiple beats
Framing + Subject + Action (beat 1, beat 2, beat 3 — sequenced verb phrases) + Setting + Camera (one move) + Light + Look + Audio (SFX per beat plus a bed — all mouths closed, nobody speaks) + Anchors.
e.g. Sequenced beats are sanctioned here specifically because official guidance that allows them pairs them with timestamps — use timestamp ranges where a model supports them.
6 — CONTINUATION-STITCH: describe only the delta
Anchors (the last frame of the previous shot, as the primary reference) + Framing (only if it changes) + Action (what happens next, from this exact position) + Camera (the move between the two states) + Audio (continues from the previous shot, plus any new event) — do not restate subject, setting, light, or look; they're inherited from the anchor.
e.g. The template's real job is suppressing re-description — re-describing appearance in a shot that's anchored to an image is a documented cause of drift.
7 — CHARACTER SHEET: produces two artefacts
Front view, side view, back view, neutral pose + Subject (name, age, build, distinguishing features, full garment description, accessory placement, posture) + Plain neutral studio backdrop + Soft even key light + Clean photorealistic reference look.
e.g. The more valuable output isn't the image — it's the identity-card text itself, pasted verbatim into every subsequent shot's Subject slot. Write it once, reuse it unchanged.
8 — LOCATION PLATE: the room with nobody in it
Static wide, the angle the scene needs + Subject: deliberately empty + Setting (materials, what's near/mid/far, one background event or none) + Camera: static locked + Light + Look + Audio (the bed this room makes).
e.g. State emptiness as positive composition — "an empty platform, white tile walls, fluorescent lights overhead" — never as "no people." Naming what you're avoiding is the trap this whole discipline exists to prevent.
9 — EDIT: the same frame, minus one thing
Reference the source image, state what to keep exactly the same (position, scale, pose, style, lighting, composition), then state the single change, then restate that placement/angle/ framing must stay identical. Canonical form: "Using this image, remove the [X]. Keep everything else exactly the same, preserving the original style, lighting, and composition."
e.g. One change per pass — this is how before/after, retouch, and removal shots are actually made, never by generating a second shot and hoping it matches the first.
What changes per model family
The skeleton above is internal and model-neutral. What actually ships to each model differs — and
dialogue syntax is the single biggest per-model trap.
Dialogue syntax — genuinely incompatible across families
Don't carry one family's tag/quote convention to another — a model that doesn't expect it either reads the syntax as literal text, or burns it into frame as a subtitle.
Read the parameter schema before you write the prompt
The schema isn't just how you call the model — it's the list of things to delete from your prose. Every parameter a model exposes is a constraint that shouldn't also cost prompt budget.
Three rules that apply to every template
1. Never leave the audio slot empty, including on silent shots. An unfilled audio slot isn't silence — it's a hallucination license; models fill it with invented speech or a generic score. 2. The negative-prompt slot is conditional. With a real field: comma-separated, under 30 words. Without one: rewrite every ban as a positive state instead. 3. A regeneration edits one slot. Diff the slot, not the whole string — that discipline matters more than whether the prompt ships as prose or as structured data.
Anti-patterns
- Two camera moves in one prompt — causes jitter; use one primary instruction.
- Tangling camera motion with subject motion — keep them as separate instructions.
- More than two interacting subjects in one generation.
- Temporal connectives ("then", "next") on models that don't actually sequence multiple actions.
- Re-describing appearance in a shot that's already anchored to a reference image.
- Empty adjectives ("beautiful", "amazing", "cinematic") — replace with the specific detail they're standing in for.
- Re-running an identical prompt and expecting variety — change the seed instead.
- Editing more than one thing per regeneration pass.
- Changing lighting or camera distance between shots of the same character — the leading cause of identity drift.