Chapter 15 · Directing performance

Yasser Ashour

2 hours ago

18 min read

Part 3 · Motion

A generated actor has no inner life to draw on. It draws what you describe, and it turns feelings into faces pulling expressions. This chapter teaches what makes a person believable on screen and how to get it from a model: one visible event for each beat, a task and a few cues, blocking that keeps bodies and eyes where they belong, listeners who work, hands that pass an object in the right order, and the timing of an Egyptian scene, where the joke lives in the held beat.

In this chapter

What makes a performance believable, and why the audience's understanding is part of it
Reducing a scene's psychology to one visible event, and keeping the intent in the note
The task and the cues, and the dictionary that turns a feeling into what a camera sees
Blocking: where bodies are and where eyes go
Listeners, handovers and two-shots
The acting around a spoken line (Chapter 22 designs the speaking shot)
Egyptian performance: register, timing and the witnessed joke
Which tool can act what, and how to repair a performance

Before you start. Chapter 9 teaches the keyframe, including the mouth state a speaking frame needs. Chapter 14 teaches the moving shot: its camera, its world and its endpoint. Chapter 5 supplies the geometry of a scene. Chapter 22 designs the speaking shot and Chapter 21 is the voice.

15.1 What this stage decides, and who decides it

Performance is what a viewer reads as a person: what someone is doing, whom they look at, what they notice, and the moment something in them changes. The director decides it, in a performance note written before the prompt. The prompt then carries only what a camera can see. This stage sits between the keyframe (Chapter 9), which fixes who is in the frame and in what state, and the moving shot (Chapter 14), which gives the clip its camera, its world and its endpoint. Speech is a further layer, designed in Chapter 22.

Five things are decided here, beat by beat. What the audience understands before the beat. What each person wants, knows and keeps back. The one visible event that changes the situation. The channels the framing can show: eyes, breath, hands, posture. And where every body and every look is. The director writes them, an assistant compiles them into the prompt, and the director judges the returned clip with the sound off, asking what changed and when.

15.2 What makes a performance believable

Top commercial directors stress one thing above all: grounding. The people on screen are relatable, our friends and families and neighbours, meeting ordinary problems. Even in an absurd set-up, the performers never acknowledge the absurdity; they react as if it were normal in their world. One casting expert puts it bluntly: "What's more important than being funny or standing out… is to be believable." The test an actor applies is "go to reality": how would I really react here?

Good directors' notes to actors are small concrete actions, not labels: not "be anxious" but "clench your hands", "eyes dart to the window when she hears a noise". They use pauses and breath, so that speech sounds unscripted. The notes that come back from casting sessions are the same few: keep it grounded, don't play it big, talk to someone specific.

The audience's understanding is part of the performance. A comic beat can fail even when the line is right, the face plausible and the cut smooth, because its consequence arrives too early, its set-up is unreadable, or the performer announces the trick. For every beat that matters, know five things before you write a word:

What the audience understands before the beat, and what is still uncertain.
What each character wants, knows and keeps back, and who is addressing whom, the camera included.
The action or line that changes the situation, the event that makes the change visible, and the response.
The relationship in time that must survive: a reveal after its cue, a listener taking in a line, an action interrupting a thought.
What may vary between takes without changing the mechanism, and what would change its meaning.

Choose the contract with the audience before the style. A credible creature can respond to a situation, as the makers of Never Say No to Panda describe: believable eyes and texture, and an uncompromising attitude instead of a friendly cartoon face. Or a performer can appear to control an impossible world, as the makers of the Old Spice campaign describe: continuous staging that made the trick feel physical. "Deadpan" or "energetic" does not tell you which contract you are in; the brief and the scene do. And keep spectacle from competing with the turn: on Lacoste's Crocodile Inside, the visual-effects company describes reducing some effects so that the background destruction would not compete with the acting.

The model renders physics, not feelings. An emotion word leaks into the face as a performed expression. Give the character something to do, something to look at, and a moment when something visible changes because of something the audience can see or hear. The makers point the same way: Seedance's guide advises replacing intense emotional language such as "fanatical" or "extremely shocked" with neutral wording (its reason is glowing eyes), and Kling's advises describing "the body position, contact point, and follow-through in plain language".

15.3 From the scene's psychology to one visible event

A performance note is the director's private account of a shot: what the character wants from the other person, what stands in the way, how they go about it, what changes their approach, the change itself, and what they carry out of the shot. A working form is want → obstacle → tactic → trigger → change → residue. It stays in the shot record (Chapter 8), because the model cannot read a note and a stated motive pulls the face towards a performed emotion. Only the note's visible result goes in the prompt: a physical task, a look, a stopped hand, a breath.

"Salma wants Mina to admit he read the letter; she keeps folding laundry instead of confronting him; he gives the wrong date; her hands stop for one beat; she resumes, more carefully" is playable. "Suspicious, sad" is not.

Figure 15.1 — The note is written in full; the prompt compiles only the playable parts. The arrows are translations: the tactic becomes a task, the trigger becomes something heard, the change becomes one stopped pair of hands.

Keep the character's core out of the prompt. A recurring character's habits (how they press a point, their rhythm, what breaks their composure) live in your notes as tendencies; the prompt carries only this scene, in the channels the shot size can show. A close-up can show the eyes and the breath; a wide cannot.

Give each beat one visible change, and name its trigger: the task that carries the beat, where the eyes go, and the moment something stops, starts or turns because of a cause the viewer sees or hears.

15.4 The task and the cues

Give the character a task, then visible cues, and give any listener a task too. Two to four cues is a working band, not a quota: add only the cues this framing can show. The cue families are gaze, blink, breath, hands, head and posture. The task matters as much as the cues, because it keeps the eyes alive: an eye with nothing to inspect goes glassy.

Never let an emotion word stand alone. A bare adjective leaks into the face as a performed expression, and the makers advise neutral wording (15.2). An adjective may stay when cues pay for it.

Weak: "She is wry and warm, correcting him."
Strong: "calm, dry, unhurried, not looking up: her eyes stay down on the tea for the whole line; the spoon keeps a slow steady stirring rhythm under the first sentence; on the last syllable she taps the spoon once on the rim of the glass and lays it on the saucer; her face does not change at all."

Prefer the body to the face, and the smallest true unit to the gesture. Reliable: stillness, a blink, a slow head turn, one hand action, a weight shift, visible breath. Fragile: bodies moving together, fast changes of expression, a face turning past three-quarter, crowds in unison.

Translate the word you feel into what the camera sees. No emotion word ships unpaid. The words "cinematic" and "epic" are never used alone: write the lens, the ratio, the composition and the grain that produce them, or the scale evidence, the figure against the mass.

The word you feel	What you write instead
tense	shoulders raised, jaw set, breath shallow, hands still, gaze fixed
exhausted	weight sunk to one hip, lids heavy, slow blink, collar loosened
joyful, celebration	arms mid-rise, mouths open unevenly, bodies at staggered phases, thrown objects mid-arc
grief	stillness amid motion, hands occupied with something small, gaze below the horizon
fear	flinch away from source, protective shoulder, whites of eyes, frozen mid-reach
determination	forward lean past base of support, jaw set, nostrils flared, gaze past frame
boredom	restrained facial musculature, idle finger rhythm, gaze mid-drift
pleading	the torso leans in; the hand lifts and hangs; the second attempt at a word quieter than the first
relief	the eyes close briefly; an open exhale; the hands unclench; the shoulders drop
suspicion	the gaze moves before the head; the weight stays back; an object stays between them

Words that turn mood into light and scale (moody, atmospheric, dreamlike, intimacy, menace) belong to the stills chapters. The channels a performance uses come from acting and casting practice:

Channel	What a camera can see
Eyes	a slow blink or heavy lids for fatigue; widening for interest or alarm; narrowing for focus; looking down; holding contact
Hands	fingers tapping on a table; fiddling with a ring or a phone; hands hidden under the table; a hand that reaches out and withdraws
Posture	a straight back; slumped shoulders; leaning in; leaning back; a shift of weight
Breath	a sharp intake; a slow exhale; quick shallow breaths; a breath taken before the line

On Kling, prefer the rows that use the body: micro-expressions are on its fragile list (15.9). For subtle acting, a recorded performance reference sent to Seedance as a video reference is a lever worth a scratch test (Chapter 13).

Template: the performance lines of a motion prompt, test grade (written for this book; not run)

[Name] [the task, as an action]. [Cue 1]; [cue 2]; [cue 3, only if the framing shows it]. When [a trigger the viewer sees or hears], [one visible change]. [The listener's task and where the eyes go.] The shot ends with [the held state].

15.5 Blocking: where bodies are and where eyes go

Blocking is where people are in the room, which way they face, and how they move. On a shot from a start frame, the frame carries it; on a shot built from references, the words must. Either way you need three notes before you write; the performance note is the one that matters most.

A geometry note for the scene: the room's fixed landmarks (table, doorway, window), who sits where relative to them, which side of the eyeline the camera keeps, and who is in frame at the start (Chapter 5).
A performance note for each shot (15.3), kept in the shot record and never pasted whole into a prompt.
A state and contact note for each prop that moves: who holds it before, during and after.

Compile only the playable parts: the task, the cues, the trigger, the contact order, the camera's one intention or one named device with its end (Chapter 14), and the ending. On return, judge: can a viewer tell what changed without reading the prompt?

Put action-critical people and props in place from the first frame. State who is already in frame. If a late entrance is meant, write it deliberately, with the doorway and who makes room.

Place people against the room, not only the screen: location, then landmark, then relative positions, then, if needed, the side of the screen. "Cal is left of the cluster nearest the red lamp." On a reverse, world positions stay and screen positions change.

Separate where the body faces from where the eyes go. Without that, the model turns the whole body for every glance. A published Seedance prompt method gives the form: "Oli's torso stays angled toward Cal on screen-left, but his eyes drop to the wet cobbles after Cal mentions Maya."

Give the shot enough seconds for its chain: notice, decide, act, consequence, and the landing and hold of any camera device (Chapter 14).

15.6 Listeners, handovers and two-shots

Let the listener work. A listener takes in what is said and begins to react during the other person's line: judging, deciding, looking for an opening. The reaction starts when the meaning lands, not when the other person stops. Carry the last shot's breath, grip or gaze into the next, and build a reverse from that retained state, never from a neutral portrait. In a group, give each person a different target and a different timing; an ensemble reacting in unison needs a reason. Give hands their own business, never gestures that mime the words. A pause is a performance, not an empty space: make it assessment, concealment or decision, or shorten it.

A handover is the hardest thing to direct in generated motion, because order matters and fingers merge. Two people in one frame: an object changes hands, or a look passes between them. You need a written ledger of who holds what before, during and after, and a decision about who acts first and who waits. On Kling the start frame is the only thing that carries identity, so use one relationship keyframe that holds both faces (16.2). When the order or the grip fails twice, change the shot: two singles and a cut, or Seedance.

Figure 15.2 — The ledger of Example 15.1, the key across the dealer's desk. Write the handover in this order; a release written before the weight is taken invites an object that floats or stays with the giver.

Write contacts as support, grip and release, in order: approach, contact, shared support, release, placement, the state after. Slow the contact, show the hands and the prop in the keyframe, and write limb, force, sequence and inertia. One instance of each prop. A reaction only after its visible cause.

Order a handover by contact: the receiver takes the weight, and only then does the giver let go. This is prop-handling craft written as a sequence. Order a look by who waits: give the responder something to keep doing until the cue. "The older looks at the box, then raises her gaze to her sister; the younger responds" lets the younger sister look first, because she has nothing to do; "keeps watching the box until the cream-clad sister looks toward her, then responds" fixes the order. Tell two people apart by something they hold or wear, and say which hand does what.

A cut between two separately generated two-shots takes its state from the first shot's accepted last frame, not from the plan: the grip, the object's pose and where the bodies are. Identity comes from the masters, not from the neighbouring shot (Chapter 10). Hand on the ledger's exit state as the next shot's entry state.

Written for this book · automotive and luxury · Kling 3.0 standard, start frame, sound off, 5 s, 16:9 · not run

Example 15.1 — the key across the dealer's desk: a handover and a look in order

The camera is locked on a medium two-shot across the dealer's desk; all the movement belongs to the two people and the showroom. The dealer holds the key fob upright on his open palm at desk height and, over one and a half seconds, lowers it about ten centimetres toward the buyer in a slow settling motion. The buyer, who has watched the fob and nothing else, slides her open hand beneath his palm and takes the fob's weight, and only then does he withdraw his hand. She closes her fingers around it and rests it on the desk edge, her thumb on its button. She keeps her eyes on the fob until the dealer lifts his eyes to her; then she raises her gaze to his, and her shoulders ease by a few degrees. One transfer, one look. In the glass wall behind them a reflected car drifts steadily from left to right across the width of the glass over the full five seconds, and the desk lamp holds steady. Keep both faces exact, the fob a single object with one button and one ring, and both hands complete, no finger merging into the other's. The clip lands on the two of them holding the look across the desk, the fob at rest under her hand, and holds there for the last second. No captions, no subtitles, no on-screen text.

Why it works: the transfer is written as support, contact and release in that order (Figure 15.2), so the weight moves before the hand does; the buyer's task (watching the fob) lasts until a cue, so the look cannot arrive first.
What to copy: the frame owns the faces, the clothes and the room, so the prompt names only what moves and what is protected; one measured magnitude for the moving object (ten centimetres over one and a half seconds) and none for the camera, which is declared fixed as a design.
The ending: a cuttable two-shot and a one-second hold; the sound switch is off in the settings, so the prompt carries only the text line.

15.7 The performance around a line

A speaking shot needs the acting to be already in the picture. Chapter 22 designs the mouth's exposure and builds the routes; what the performance around the words needs is short.

The acting is the ceiling. On the exact-take route a lip-sync pass redraws the mouth and the face around it; it cannot add head, eye or breath performance. On the new-performance route the model acts the line from your prompt. Either way the direction goes into the picture prompt (22.4, 22.6).
Business first, then the line. Seedance documents relative time and ordered beats ("After 3 seconds …"), so a beat before the line is a design choice. Write the listening as business, a breath or a blink, not as stillness: something visible moves from the first second.
Write the mouth and the gaze. "Lips closed and relaxed", "eyes stay on the glass" and "lips fully visible" are direction, not emotion.
Land the line. After the last word the mouth closes, the breath releases and the gaze holds.
One speaking mouth in the frame, and nothing crossing it: no hand, cup, smoke, veil, microphone or prop. Make the other person an edge with a job: a shoulder, a hand, a spoon.
Give each department its part. The voice carries the objective, the breath and the stress; the picture carries the face and the body; the edit carries the exact moment. When a gesture must hit a word, decide whether the route can hold it or the cut must.
Cover a difficult word with the cut, as directors do in live action with dubbed speech: a cutaway to the listener, an insert on the difficult word, the wide for lines where mouths are small (the exposure tags of 22.3).

15.8 Egyptian performance: register, timing and the witnessed joke

Casting is by behaviour, not by type. Define a role by energy, goals and temperament, and by quirks ("constantly reorganises the pen holder when anxious") rather than one-word tags. Culture and class are written as behaviour, never as a stereotype: "back straight and not touching the chair" says a retired teacher without a word about class. Where a gesture is cultural (a greeting, a hand on the heart), check it with someone who knows it.

Egyptian comedy is set-up, held beat, comment. The held beat is the joke: "her face does not change after the tap". In a dialogue comedy the hold is usually carried by the listener's face, and the laugh forms in the pause before the comment (the تعليقة).

The joke is witnessed. In Egyptian comedy the reaction shot (the wife's look, the colleagues' frozen faces) is half the laugh. This is exactly where generated scenes fail first: one face speaks acceptably, and the faces listening do nothing specific. Direct the listener as carefully as the speaker, with a task, a target and a timing of their own.

Understatement in the victim, excess in the perpetrator. Egyptian audiences reward deadpan and speed more than mugging.

Register is cast, not written. Colloquial Cairene is the default register of Egyptian advertising; Modern Standard Arabic reads as broadcast-formal and flattens comedy. Within Cairene there are sub-registers (the kiosk and the microbus, the upper-middle Cairo that switches into English, Upper Egyptian, Alexandrian), and the vocatives (يا أستاذ، يا عم، يا حبيبي) tell a native ear the relationship and the class at once. Real conversation overlaps and is full of particles (يعني، بقى، طب); synthetic dialogue takes turns and sounds like a table read. Two speakers in one frame are not yet a reliable route, so the default for overlap is singles, joined in the cut and the mix (Chapters 22 and 29).

Written for this book · telecom and tech · an Egyptian family, Seedance 2.5, references mode with a start frame, sound off, 720p, 5 s, 16:9 · not run

Example 15.2 — the data bill: three listeners, three timings, one held face

Image 1 is the first frame, exactly as it is. At the dinner table the father holds up his phone to show the month's data bill, and the family watches. The camera is locked at the end of the table, a medium-wide frame; all the movement belongs to the family. 0-2 s: the father tilts the phone toward the table and sweeps his thumb down the screen, satisfied, saying nothing; the grandmother keeps peeling an orange and does not look up; the daughter's eyes go to the phone and then to her brother. 2-3.5 s: the son, whose spoon is halfway to his mouth, stops; the spoon stays in the air and his face does not change, while the father turns the phone a few degrees further toward him. 3.5-5 s: the grandmother drops an orange segment into her mouth and looks at the son at last; the daughter lowers her eyes to her plate and presses her lips together. Nobody speaks, and nobody hurries: no sudden gesture, and the son's stillness lasts from 2 s to the end. Keep all four faces exact, and the phone screen a soft glow with no legible text. The clip lands on the son with the spoon still raised and the grandmother's eyes on him, and holds for the last half second. Steam lifts from the rice and a ceiling fan turns slowly. No subtitles, captions or on-screen text.

Why it works: each listener has a different target and a different timing (the grandmother a task, the daughter a chain of looks, the son a held beat), so the ensemble does not react in unison; the joke is the son's unchanging face, written as "his face does not change".
What to copy: ordered beats with whole-second stamps, the form the maker documents for Seedance; the subject and event first, then one camera instruction; people protected by name after the action; a landing on the held face.

Judge on a scale agreed in advance. Three levels calibrate an Egyptian dialogue scene before any AI output is judged. Weak: the register or the rate is wrong, the pitch is even, the dialogue takes turns, the lips only approximate; the ear rejects it within seconds. Acceptable: native speech at the right pace, lines that land, but listeners who are passive; charm carries it. Aspirational: the register cast to the character, the two-beat timing, overlapping particles, listeners who react specifically, lips and jaw matching in close-up.

15.9 Which tool for which job

Job	Tool	What it can act	Section
A performed face in a still	Nano Banana Pro, GPT Image 2.5	adjectives produce performed faces; state the muscles instead ("mouth closed and relaxed, brows level")	11.3, 12.3
A silent performance from a start frame	Kling 3.0	reliable: stillness, a blink, a head turn, lifting a cup, breath; fragile: micro-expressions, precise eyelines, hands on small objects, turns past about sixty degrees (other directors report)	16.3
Ordered beats, an ensemble, a speaking performance	Seedance 2.5	the maker documents ordered beats and relative time	17.3
An economy performance	Wan 3.0	not yet known as an actor	18.3
A face that speaks an approved take	Sync Lipsync 3	no prompt: the talking plate's acting is the ceiling	24.3
Speech from the model itself, or a sound-led beat	Veo 3.1, MiniMax H3	not yet known as actors	19.3

Kling 4.0 was announced on 28 September 2026 and is due in October; until it ships, performance is directed on 3.0. Re-read 16.10 when it does.

15.10 Repairing a performance, and the gate

When a performance fails, change the direction, not the adjectives.

You see	Change the direction
The actor advertises the emotion	give a practical task and cut the feeling word
The outcome shows before its trigger	limit the character to what they know at that moment; move the trigger earlier in the text
The listener waits mechanically	let understanding begin during the other line
The gestures illustrate every word	give the hands their own business
An empty pause	make it assessment, concealment or decision; otherwise shorten it
Every cut resets the actor	carry the previous action and its residue (Chapter 29)
The ensemble moves in unison	a different target and timing for each person
The whole body turns for every glance	separate the torso from the gaze
The wrong person reacts first	the responder keeps a task until the cue
The prop stays with the giver, or floats	"takes its weight, and only then … lets go"; the giver empty-handed after
Identities merge in a two-shot	simplify reflections and crowding; props that tell them apart; singles
A mouth opens uninvited on a silent clip	write the mouth: lips together and relaxed; give the listener no line that invites speech

Stop after two. Two wording repairs of the same performance, or two failures of the same blocking, and you change the input: a better frame, a layout sketch, a blocking clip, a correct contact frame, or two singles and a cut. Follow the retry ladder of 14.16. Record with the verdict whether the wording was task-led, cue-led or both, so the house learns which works.

The acceptance check for a performance

Watched with the sound off: a viewer can say what changed, and when, without the prompt.
Every emotion word is paid for by a cue; no face acts a feeling.
The listener has a task and reacts when the meaning lands, not when the line ends.
Hands and props follow the ledger: weight taken before release, one instance of each prop.
The look order is right, and the last frame is a cuttable composition.
The exit state (grip, gaze, breath) is written down as the next shot's entry state.

The records this stage writes. In the shot record: the performance note (want, obstacle, tactic, trigger, change, residue), the geometry note for the scene, the state and contact ledger for every moving prop, and the exit state. In the run record: the verdict on the returned clip, with the wording style (task-led, cue-led or both). The gate is acceptance: the director's eye, on the clip with the sound off, then again in the running cut (Chapter 29).

Open questions

A recorded performance reference for subtle acting, sent to Seedance as a video reference: a lever worth a scratch test on the first project that needs it.
Two speakers in one frame, on any route: prove it on a scratch clip before a client shot depends on it (Chapter 22).
Wan 3.0, Veo 3.1 and MiniMax H3 as actors: no reliable reading yet; test on a scratch clip before a client shot depends on them.

What to remember

Believable means grounded: relatable people, ordinary problems, no winks at the absurd, small concrete actions.
Know what the audience understands before each beat, and which relationship in time must survive.
Want, obstacle, tactic, trigger, change and residue stay in your notes; the prompt carries the task, a few cues, the trigger and the one visible change.
Never let an emotion word stand alone: translate it with the dictionary of feeling into physics, and prefer the body to the face.
Separate where the body faces from where the eyes go. A handover: the receiver takes the weight, and only then does the giver let go.
Around a line: business first, then the line, then a landing; one speaking mouth, and a listener with a job.
Egyptian comedy is set-up, held beat, comment, and the joke is witnessed: direct the listener as carefully as the speaker.
Two failed repairs, then change the input, not the adjectives.