Cinematic prompts · Part 4 of 5
Continuity is cinematography
Lock the face, the coat, the room, the lens height, and the light. The move is what you are allowed to change.
The camera cannot tell a story if the face, the coat, and the room reset every cut. Pretty singles are easy. A sequence is a lock. Treat identity, set, lens height, and light as the shot. The move is what you are allowed to change after those four are frozen.
What is a lock?
A lock is a fact the next prompt is not allowed to reinvent.
- Face and body as a reference image, not a paragraph alone
- Outfit in the same sentence every time
- Set as its own reference. Match it. Do not redraw it.
- Camera grammar (lens height, 180 line, direction of travel)
- Light direction so a push-in does not relight the scene
If any of those drift, you do not have coverage. You have a new movie that happens to share a title.
The Godfather — the coat is the sequence
Michael’s look does not reset between the wedding, the hospital, and the restaurant. Same man. Same gravity. New camera job.
Goodfellas — the club is one space
The Copa walk only works if the kitchen and the dining room belong to the same night. The track reveals belonging because the rooms do not rebuild themselves.
AI fails the second film more often than the first. Faces swap. Wallpaper mutates. Windows jump sides.
What to start with
- One hero still you will reuse (Image 1 is this person)
- One empty-room still if the scene stays put (Image 2 is this set)
- The outfit sentence, copied exactly
- Lens height and which way is camera left
- The Director’s sentence for this beat only
Do not start with a new wardrobe paragraph “for variety.” Variety is how sequences die.
Lock four things before you lock the move
1. Character.
Bind the still. [Image 1] is the detective. Same face, same body, same navy coat. A prose description of “a tired man in his 40s” will mint a cousin every take.
2. Set.
Bind the room. [Image 2] is the alley. Match it exactly. Do not redraw the set. Empty-room references beat hoping the bricks remember themselves. Clark’s rule for text-only: restate the visual anchors in every prompt. Do not write “still,” “previously,” or “continues.” The model has no memory. Clark
3. Camera grammar.
Pick a lens height and a line and keep them. If shot one looks screen-left toward the street, shot two cannot invent a new street behind his head. One grammar per scene. Plaza
4. Light.
Name the key. Key from camera left, neon behind him, wet bounce on the coat. A push-in that also changes noon into night is not a push-in. It is a recast.
Kapwing’s four layers exist so you can swap only layer four.
- Bind the reference
- Scene in one sentence
- Look line identical every time (
35mm anamorphic, shallow DOF, wet neon) - The move, copied from the library
Two ways to build the lock
From a still you trust.
Find a frame with the composition you want. Ask a chat model to describe scene, framing, lens, and light. Use that block as the look line in every later prompt, or as frame one if you own the image. Kōda
From a first AD.
Plaza’s custom GPT is built for this job: composition, camera motion, character lock, lens and light consistency, tone, even ambient audio. Use it to pack the four locks before you touch Veo or Seedance. Plaza
PJ Ace’s Odyssey pipeline makes the same point from the other side. Video prompts cannot rescue a bad first frame. If the still is wrong, every second after it is a fight with that mistake. Imagine thread · PJ Ace
Jay’s production version of the same lock: sort faces, wardrobe, locations, and props before you generate. Specify the accent. Grey backgrounds on character sheets. Fewer bodies in the frame. Jay
For faces, newer Seedance work splits the sheet apart. Give the face its own tight portrait, chin up, with no collar to compete with the outfit. When the shot turns the head, add a three-quarter view and a profile as separate images made from that portrait. A sheet of views is often read as several people, and repeated front shots get averaged rather than learned from. Send only the angles the shot shows. Starrd · Atlas Cloud
Worked swap (same alley, only the camera changes)
Engine: the detective in the navy coat, alley rain, cigarette.
Shared lock (copy into both prompts)[Image 1] is the detective. Same face, same navy coat, wet shoulders. [Image 2] is the alley. Match the bricks and the dumpster. Do not redraw the set. Key from camera left, neon behind him. 35mm, shallow DOF.
Shot A — pressureHe cups the match and lights the cigarette. Medium close-up, eye level. Locked-off static shot. Camera fixed on a tripod. The stillness reveals he is waiting.
Shot B — accusationHe cups the match and lights the cigarette. Medium close-up, low angle. Slow dolly push-in over 4 seconds. Camera physically moves forward. No lens zoom. The push-in reveals the tension in his face.
Point at what changed: size or angle or move. If shot B grows a new scarf, a dry coat, or a second alley, the lock failed. Regenerate B. Do not “fix” it by rewriting the world.
Kill test
Run this before Generate.
- Same face still bound?
- Outfit sentence identical to shot one?
- Room reference bound, or the room described with the same nouns?
- Same lens height and key side?
- One named move?
- End frame described so the next shot has a landing pad?
- Any word like “still,” “previously,” or “continues”? Cut those.
If the coat changed and the move was perfect, you do not have a shot. You have a recast.
Using AI as a writing partner
Use chat to pack the lock. Use the video model only after the lock reads as copy-paste.
Paste into your AI chat
When two clips share a face, a coat, a wall, and a key, the move finally has a story to tell. Next lesson is how the best people block the ambitious shots that sit on top of this lock.