Video Prompting

Video Prompting

Write a production brief, not a vibe paragraph. Models fail when every asset competes for identity, motion, camera, and style.

Core rule

Assign authority before action. Every reference gets one job and an explicit boundary on what it must not control. Then: lock continuity → place screen geography → time the shot → light/sound → short negatives.

When to use which skill

Session gates

1. get_timeline — if canGenerate is false, ask user to sign in / subscribe.
2. list_models for image / video / audio — ids and caps drift.
3. Propose prompt + model + duration + aspect (+ refs). Wait for confirmation — paid, not undoable.
4. Fire generate_*; do not busy-poll. Ready when get_media({ ids }) has no generationStatus. On failed, report and ask before retry.
5. Organize with folder (e.g. Hero/Sheets, Hero/Takes).

Reference wiring

Declare authority when multiple refs are passed:

@Image1 defines identity and wardrobe only. Ignore its background.
@Video1 defines motion timing and camera only. Do not inherit its people or location.
@Audio1 defines <Character>'s voice timbre and delivery only.

Never: “use all references to make the same scene.”

Prompt architecture

Stable truth first; time second; polish last.

[REFERENCE USE]     — what each @ImageN / @VideoN / @AudioN controls + ignores
[IDENTITY LOCKS]    — count, face, hair, wardrobe, props, voice ownership
[SCENE]             — place, time of day, dramatic goal, emotional turn
[DIALOGUE]          — exact quoted lines, speakers, delivery + voice pins
[SCREEN GEOGRAPHY]  — who/what is left/right/front before anyone moves
[SHOT LIST]         — time ranges: framing, action, dialogue, end state
[ACTING]            — posture, gaze, gesture, emotional progression
[LIGHT / IMAGE]     — light, palette, texture, lens, DoF
[CAMERA]            — sizes, move rules, axis, transitions
[SOUND]             — ambience, dialogue clarity, music rule
[NEGATIVES]         — only the failure modes that matter

Length vs start frame: with startFrameMediaRef alone and simple motion, keep the prompt short (camera + action + speech/SFX) — do not re-describe the still. With multi-ref / timed performance / edits, use the full architecture.

Continuity + shot timing

Lock subject count, face/hair/wardrobe, prop ownership, screen direction, voice ownership. Place people before they move.

Treat duration as an editorial budget (list_models → durations). One primary beat per range: establish → escalate → resolve. Prefer one beat per generate_video call; assemble on the timeline. If dialogue rushes, shorten lines — do not cram two turns into one short clip.

Model routing

Confirm ids/caps via list_models before firing. Default stack for these skills:

Flux 3 draft → enhance (not upscale)

1. Iterate: generate_video({ model: "flux-3", draft: true, … })
2. When get_media shows the draft ready and canEnhanceDraft: true, final: generate_video({ enhanceDraftMediaRef: "<draftId>" }) alone — no prompt/model/inputs. Keeps motion; renders FLUX.3 at 1080p.
3. Do not use upscale_media for this path. Enhance only works on Flux 3 drafts (canEnhanceDraft); other models have no enhance step.

Precise editing

1. Master — source / @Video1 owns timeline, subjects, camera, audio.
2. Changes — exact subject/object/region/line/time only.
3. Preserve — identity, occlusion, eyelines, camera, dialogue, ambience, duration.

Text and titles

Most video models cannot render reliable readable text — use generate_image plates or add_texts. Exception: MiniMax H3 / H3 Max kinetic flat type with an exact string allowlist — follow motion-graphics (do not improvise long paragraphs of type).

Recipes (near-look patterns)

Beat-synced stylized title

Silhouette motion via generate_video; title plates via generate_image; cut on music with detect_beats.

@Image1 = silhouette identity only. Ignore photoreal BG.
Flat black silhouettes on solid poster colors (2–3 saturated fields). No naturalistic shading.
Comic-panel crops; freeze-holds dead-still for one full beat; hard cuts only; heavy grain/gate weave.
Alternate kinetic bursts with freeze holds. End frozen mid-pose on a solid field.

Then: generate geometric title stills → detect_beats → place clips/stills on downbeats.

Texture morph

@Image1 = source surface texture only. @Image2 = destination landscape only.
Continuous macro push; transform when contours align. No hard cut, tear, or seam.

Motion transfer

@Image1 = identity/wardrobe only. @Video1 = motion/timing/camera only — not its cast or location.
Ordered action with stable L/R geography. Physically believable contacts and shadows.

Voice lock

Reuse pitch + tone + accent + pace verbatim every beat. Better on Seedance: pass beat-1 audio as referenceAudioMediaRefs (@Audio1 = voice only) on later beats.

Surgical edit

Source is sole master. Replace/remove/add only listed items.
Preserve identity, timing, occlusion, eyeline, camera, layout, dialogue, ambience. Nothing else.

Troubleshooting

Debug order: authority → locks → timing → preserve → then adjectives.

Bakeoff notes (T2V, cheap 480p/720p, Aug 2026)

Project: Skill Bakeoff CM+VP · timeline Video Prompting Bakeoff. No refs — text-only stress test of architecture vs vibe.

Checklist before generate_video

• User confirmed cost; list_models current
• Every ref has authority + ignore; tags match tool arg order
• Subject count + geography locked
• Duration fits one beat; dialogue fits budget
• No readable text expected from the video model
• If editing: master + changes + preserve