Film Direction (stage 0)

Why a separate stage

The viewer of a film cannot say why it is good or bad. They feel pace, tension, and continuity, but they cannot name them. A product trailer viewer is the same: they either get the "aha — so that's what it does" moment or they don't. Everything that produces that feeling is a decision the director made before shooting. AI generation does not change this; it only makes the cost of undecided things visible faster.

So before a script exists, write down the decisions. The output is films/<slug>/direction.md.

1. Intent (write it in three lines)

• What kind of film: drama short / product trailer / music-video / documentary-style piece.
• What the viewer must feel or understand by the end. For drama: one emotional sentence ("a man forgives himself a little"). For a product film: the list of "aha" moments in the order they must land ("it understands a plain description → it builds a world → you step inside").
• Length and unit: total runtime; block length (15 s default because one Seedance generation = one block).

1b. The dramatic spine — see film-story/SKILL.md (stage 0b), which owns the full engine: spine, conflict ladder, loss, dilemma, setups/payoffs, emotion map, nested stories. Run it before the script. Summary of the minimum here so this file stays self-contained:

Every block must then carry a tension line: what is unresolved at the end of this block that makes the viewer need the next one. A block whose tension line is "nothing — it went well" is a montage, and a film may afford one at most.

Product films are not exempt. The product is never the hero; a person is. The product enters as the thing that changes what is possible for the hero against their obstacle — and it should arrive after the obstacle is felt, not before. If the product removes the obstacle instantly there is no film; the product should open a door the hero is afraid to walk through.

2. Genre → defaults

Genre is not decoration; it sets defaults for everything downstream, so name it once here.

3. The quality function — variables that decide a shot

Treat a generated shot as quality = f(x₁ … xₙ). The course discovered these variables one failure at a time; list them here so every prompt sets all of them on purpose. If a variable is left unset, the model chooses — and the model's choice is the "slop".

A. Reference / identity variables

• Face identity (sheet), hair, facial hair state (stubble vs clean-shaven per scene), skin texture level.
• Wardrobe per scene (kit on the pitch, casual in the office; the same outfit in both would be wrong unless the story explains it).
• Recurring props (ball, notebook, trophy) — their own sheet.
• Featured one-off characters (the old fan) — described in text plus an identity-isolation clause so the hero's face does not leak onto them.
• Location plate — used as look/scale/layout reference only, with the statement that it is not a keyframe.
• Screen content (a laptop showing a real product page) — a real screenshot, never described from memory.

B. Spatial variables

• Counts: people, chairs, balls. Unlocked counts get filled.
• Seating / positions in metres, screen-left/right, who faces whom, the 180° line.
• A wide anchor at the head of any multi-shot block in a known room.
• Edge placement and looking room (or deliberate absence of it).

C. Camera variables (per shot)

• Lens (24/28/35/50/85/100 mm or 8° super-tele), height, angle (level / Dutch °), move (static / push cm / whip with timings / rack / crane).
• Camera operator character: locked-off, gimbal, handheld 1–2 cm, super-handheld 3–6 cm with breath and whip-lag.
• Shot size must not change inside a single continuous move (a whip pan opens and lands on the same size).
• Frame-rate honesty: real time vs an explicit speed ramp with start/end beats.

D. Performance variables

• Beats timed to the half-second; muscle-level actions (masseter, swallow, inner brow, chest rise, foot tap rate).
• Eye-line target and eye-tracking stability (drifting eyes were the most common reject in the final scene).
• Emotional entry state — must match the previous block's exit state (no smile two seconds after tears).
• Dialogue: exact words, one speaker per shot, tone notes (micro-pauses, trembling, self-resentment), where the voice breaks.
• Anatomy when a body part fills the frame (hands: veins, tendons, knuckle creases, hangnail, uneven nails).

E. Light / atmosphere / palette

• Single stated key source and direction, WB in Kelvin, contre-jour or wrap, haze %, fog, flare style.
• 60:30:10 palette named in colours, contrasting between intercut locations (warm office vs cold stadium vs golden street).

F. Time / pace

• Shot count per 15 s (2–3 normal; 6 only for action montage with salvage expected).
• Held pauses written as beats, especially 1–2 s of processing after any flashback before dialogue resumes.
• Terminal action inside the last beat ("the shot cuts as he leaves") so takes do not end early.

G. Sound

• Diegetic list; the one line; "no music". VO lines recorded as a separate face take when the action take is already locked.

H. Text on screen

• Exact words, quoted; hold ≥2 s on the surface; camera angle that makes it legible (high angle on a notebook).

Copy this list into direction.md and tick which variables matter for this film. Prompts later cite it.

4. The shot-justification test

Every shot in the script must answer, in one line each:

1. What does it reveal (plot fact, emotion, product function, spatial fact)?
2. What breaks if it is cut? If nothing breaks → cut it. (The course cut "shot 4: he turns and says 'And I missed'" because the unfinished "And I…" in shot 3 already carried it.)
3. Why this size and angle rather than the default? (A Dutch angle because he is losing grip; an OTS because the question is aimed at him; a worm's-eye rise because the boy's world starts at the ball.)
4. What must the viewer be able to see for the shot to work (hands textured, text legible, ball in frame)? → these become asset and constraint lines.

For product films add: Which "aha" does this shot deliver, and can a first-time viewer perceive it without narration? If the answer is "they wouldn't notice", redesign the shot (bigger screen, slower move, a hand pointing) or cut it.

5. Shot-to-shot logic (the part the course felt but never wrote down)

The author of the course kept saying things like "any director will tell you…" — these are the rules he was applying:

• Question → reaction: a line aimed across a room is followed by the face it lands on. Cut after the line, not during.
• Flashback → processing pause: never cut from a memory straight into dialogue. Give 1–2 s of face (reuse an existing clean take if one exists).
• Anchor → detail: wide first, then hands/face. The wide is what keeps the room consistent across cuts.
• Escalation through camera, not just acting: when the emotion is right but the shot is flat, change lens, angle or operator behaviour; do not push the actor harder.
• Tension tools: Dutch tilt, no looking room, tighter lens, less headroom, handheld tremor. Release tools: level horizon, wider lens, more air, static.
• One continuous take for a physical climax (the penalty) so the physics stay honest; cut coverage for dialogue so pauses can be placed.
• Speed ramp only inside an already-locked continuous take, and only around one named beat.
• Whip pan = one move, same shot size both ends, labelled subjects, timed swing.
• End on the beat, not after it — the penalty take ends on the save, not on a reaction; the reaction gets its own shot where the pause can be controlled.
• Silence is a beat. Write it with a duration.

6. Palette per location

Write a table: location → time of day → key light → 60:30:10 → mood word. Intercut locations must contrast (warm/cold, or warm/golden). This is what makes an edit read as a flashback without a title card.

7. Continuity ledger (continuity.md)

For every block, on lock, record the exit state:

| block | character | seat/position | wardrobe | props in hand | emotional exit | eye-line | last-frame file |
|---|---|---|---|---|---|---|---|
| 1 | santiago | lounge chair, upper-right | camp shirt | none | eyes closed, holding it in | down | scenes/block-01-last.png |

The next block's prompt must open on these states. This single table prevents the "he smiles right after crying" class of failure and the "different chair" class.

8. Output

direction.md with sections 1–7 filled. Show the user the intent, genre defaults and palette table; ask for corrections; then hand off to film-script. Do not generate anything in this stage.