Drama director

Design the scene; creative-shortdrama renders it. Splitting them matters because the first capability test produced eight technically clean shots with no dramatic shape — the pipeline was never the problem.

Read references/scene-and-acting.md before directing anything. It holds the craft: beat/turn/button, coverage for singles-only dialogue, the space-map contract, the performance vocabulary that survives the model, and what each reference line contributes. This file is the procedure; that one is the reasoning.

Mode 1 — find the turn (do this before any shot list)

Given a premise or a script, answer four questions in writing. If any answer is weak, the shots cannot save it.

1. What does each character want in this scene? Not in the story — here.
2. What is different at the end? If the only answer is "the audience knows more", there is no turn. Fix the scene, not the coverage.
3. Where does the turn land? Name the shot. In a two-hander it is almost always the listener's face.
4. What is the button? The last image. Usually silent, usually after the last line.

Then break it into 3–6 beats, one line each: who attempts what. A beat is an attempt, not a topic.

Refuse to proceed on a scene with no turn. Say so plainly and propose the cheapest one that fits the material — see the turn table in the reference.

Mode 2 — the shot plan

Produce a table, one row per shot. These fields, in this order:

Field	Rule
shot	S1, S2…
narrative_function	setup / attempt / turn / reaction / button — every scene has exactly one turn row
frame	CU single / MS single / wide (silent only) / insert
who	Character, or "none" for an insert
screen_side + eyeline	Quoted from the space map, never invented per shot
line	The dialogue, or "—"
performance	One micro-event in physical vocabulary: eyes down, blink, jaw sets, hand stops
stillness	What must not move
mood	One or two words. Varied down the column — if every row says "quiet", the scene is flat
motion_amplitude	micro / small / medium. Dialogue defaults to micro
motion_floor	subtle / visible
gesture_state	full / reduced / suppressed, for the recurring gesture (or blank)
seconds	Intended length

Before the table, write the space map once: camera position, who stands screen-left and screen-right, where the door and other landmarks sit. Every screen_side/eyeline cell quotes it. Reverse angles come from the map, not from what looks good in a given still.

Coverage floor for a two-hander — the minimum that cuts: orientation (silent) · single A · single B · reaction · insert · button.

Mode 3 — critique a scene or a cut

Run these checks, in order, and report failures with the shot named:

1. Turn — name the shot where it lands. Missing ⇒ the scene is an exchange.
2. Button — is the last beat an image, or a line trailing off?
3. Reaction protected — does the cut reach the listener before the speaker finishes? A cut on the last syllable is the flattest option.
4. Mood variety — same mood every row is a flat scene.
5. Performance — is each shot's direction one physical event, or an emotion word the model will answer generically?
6. Stillness stated — is what must not move written down?
7. Space map honoured — eyelines and screen sides consistent, inserts on the correct side of the line.
8. Action beats — one verb per shot, object already in contact at frame 0, destination named, no cut inside the action.
9. Renderability — no camera move, no dialogue in a wide, no shot asking for speech and body motion at once, nothing over ~8 s with large motion.

Working rules

• Write what the face does, never what it feels. "Eyes drop, then back to him" renders; "conflicted" does not.
• Subtext is a withheld action. "His hand starts toward the cup and stops."
• One gesture across the scene, diminishing. Decide it here, in Mode 1 — it is the arc the face cannot carry alone.
• The insert is structural. People-free, same space, one motion with a stated physical cause, placed after a beat.
• Design around the renderer, not against it. Static frames, singles, stillness broken by one event — that is the reference lines' language anyway.

Hand-off

Give creative-shortdrama the space map and the shot table. It owns stills, TTS, i2v/v2v, the measurement bars and the assembly. Its framing rules — CU for dialogue, two-handers as singles, beats over ~1 s of speech — are constraints on this plan, so read them before writing the table rather than after.

Cross-references

• references/scene-and-acting.md — the craft, and the four reference lines
• creative-shortdrama — the renderer
• creative-mv-art-director/references/montage-and-mv-editing.md — cutting, entry relations, the shot-state ledger
• {vault}/70_Media/references/scene-acting/ — reference scenes; measurements turn the reference's §9 questions into numeric targets