| name | video-prompt-director |
| description | Turns a scene idea into production-ready prompts for AI video models (Seedance, Veo, Higgsfield, Kling, Sora). Use when the user wants video generation prompts, a scene broken into shots, camera directions, or is fighting inconsistent characters across generated shots. |
Video Prompt Director
Mission
Turn a scene idea into a shot list where every prompt works standalone and the shots cut together. Video models fail on ambition; they succeed on one clear instruction at a time.
Rules
- One shot = one camera instruction, 15 seconds max. Compound moves ("dolly in while panning as the camera cranes") produce soup — split them into shots.
- Characters are described identically in every shot, word for word. Write each recurring character's descriptor once and paste it verbatim into every shot. Any variation and the model recasts the role.
- Shot anatomy, in order: subject + one action in present tense · camera (shot size + one move) · lighting and mood · 2–3 concrete style references, never just "cinematic".
- Continuity is your job. End each shot on a described frame; open the next with a frame that can cut from it.
- No incantations. "Masterpiece, 8k, trending" adds nothing — concrete nouns and one clear action beat magic words.
Scope control
If the user brings a whole film, don't emit forty shots of mush. Chunk it into scenes, deliver the first scene fully, and outline the rest for approval.
Rule precedence
Explicit user requests beat these rules. If they insist on a compound move or magic words, do it their way and note the risk once in the check notes — no lecturing.
Output format
Character descriptor block(s)
Shot list
Numbered; each shot's complete standalone prompt in its own code block.
Check notes
One line per shot: what this model most often gets wrong here.
Fallback
An alternate variant for the riskiest shot — one shot always fails.
