Directing an AI actor across a long take means giving the performance somewhere to travel, because there is no cut available to rescue a beat that does not land. Short generations forgave a lot. If a five second clip held one expression, that was arguably correct. A thirty second take holding one expression is a mannequin. The performance has to start somewhere, change in response to something, and end somewhere else.
The method is the same one stage and screen directors use. You direct intention and action, not emotion.
Why Emotion Words Underperform
Asking for a sad character produces the visual shorthand for sadness, applied evenly and immediately from the first frame. It arrives without cause, holds without change, and reads as an expression rather than a state. Worse, it removes the arc, because a character who is already at the emotional destination has nowhere to go for the remaining twenty seconds.
Intention works differently. A character who is trying to leave without being noticed will produce restraint, glances, controlled movement and tension, none of which you had to name. The behaviour generates the emotion instead of the other way around.
Write What the Character Wants and What Blocks It
Two clauses carry most performance direction. What is she trying to do, and what is stopping her.
She wants to get the file out of the drawer before he comes back. She has to do it without making noise. That is a complete performance brief, and it will produce a specific, watchable behaviour set. Compare it with anxious, nervous, tense, which produces a face.
Obstacles are what create the microexpressions viewers read as acting. A performance with no obstacle has no texture.
Break the Take Into Performance Beats
Just as action needs timing, so does performance. A long take should carry two or three distinct performance states with visible transitions between them.
A typical microdrama shape: composed, then destabilised by an event, then a decision. The transitions are the whole point. Direct them explicitly. She holds her expression for two seconds after she reads it, then it goes. The delay is the acting.
Eyeline Is the Most Underused Instruction
Where a character looks and when they look away carries more meaning per word than almost anything else you can write. Looking away before answering reads as a lie. Looking directly and holding reads as a challenge. Looking down and then up reads as a decision being made.
In vertical framing this is magnified, because the face occupies a much larger share of the frame than it would in landscape. Eyeline direction that would be subtle in a wide cinema shot becomes the primary storytelling channel in microdrama.
Direct Stillness Deliberately
Undirected models fidget. They generate continuous small motion to fill time, which reads as restlessness regardless of what the scene needs. If a character should be still, say so, and say for how long. She does not move until he speaks is a real instruction, and it produces one of the strongest contrasts available when the movement finally comes.
Worked Example: A Confrontation That Was Not Working
The scene is a two hander where a woman confronts her business partner about a missing payment. The original direction was: she is furious and confronts him angrily, he is defensive and guilty.
Every generation produced two people shouting from the first frame with nowhere to escalate. The rewrite removed every emotion word:
She wants him to admit it without her having to accuse him. She stays seated and speaks quietly. He wants the conversation to end. He answers her first question, then looks at the door before answering the second. She notices him look at the door. She stops speaking and waits. He fills the silence.
Nothing in that brief names an emotion, and the result reads as a far angrier scene than the version that asked for anger. It also gives both actors an arc, and it produces a natural button when he fills the silence, which is exactly where the episode should cut.
Common Mistakes Directing AI Performance
Front loading the emotional peak. If the take opens at maximum intensity, the remaining time has nowhere to go.
Stacking emotion adjectives. Sad, devastated, heartbroken and tearful in one prompt produce an average, not an intensity.
Directing both characters identically. Two characters with the same objective produce a scene with no friction.
Forgetting the listener. The character not speaking is doing half the work. Give them something to be doing.
Leaving eyelines unstated. The model will choose, and it usually chooses camera, which breaks the scene.
Never directing stillness. Constant motion flattens every beat into the same register.
The Listener Carries Half the Scene
In a two hander, the character who is not speaking is doing as much work as the one who is, and they are almost never directed. Undirected, a model gives the listener a neutral hold or a generic reactive nod, and the scene flattens.
Give the listener an objective of their own, and make it different from the speaker objective. He is waiting for her to finish so he can leave. She is watching to see whether he looks at the door. Two objectives in tension produce a scene. One objective and a listening face produces a monologue with a witness.
Physicality as Performance
Emotional states show in the body long before they show in the face, and body direction is far more reliable to prompt than facial direction. Where the weight sits, whether the hands are occupied, how much space a character takes up, whether they face the exit. All of these are directable and all of them read.
She keeps her weight on the foot nearest the door produces a legible reluctance that no facial instruction would reliably deliver. It also survives generation better, because gross body position is more stable than fine facial nuance.
Vertical Framing and Performance Scale
Microdrama frames faces large. A gesture that would be readable in a landscape wide will exit the frame entirely in nine by sixteen, and an expression that would be subtle at cinema scale becomes the dominant element. Performance direction has to be scaled accordingly.
In practice that means smaller physical choices and more eyeline work, and it means blocking that keeps hands and props near the face if they need to be seen at all. It also means stillness reads more strongly, because there is less frame for motion to dissipate into.
Consistency of Performance Across a Season
Characters need a consistent behavioural signature across sixty episodes, or they read as different people wearing the same face. That signature is a small set of fixed traits written down once: how they occupy space, what they do with their hands, whether they hold eye contact, how quickly they answer.
Kept in the season bible alongside the visual references and pasted verbatim into each take, those traits do for performance what a reference image does for appearance. Rewriting them per episode, which is what most productions do by default, is why AI actors so often feel inconsistent even when they look identical. As of 2026 no model holds a performance identity across sessions on its own, so it has to be held in the production system instead.
Direct the Transition, Not the States
Performance quality in a long take is concentrated almost entirely in the moments between states, and those moments are what prompts leave out. Two states with no described transition produce a cut style jump in the middle of a continuous shot, which is the most common reason a technically clean take still feels wrong.
Write the transition explicitly and give it a duration. Her expression does not change for two seconds after she hears it, then it goes all at once. He starts to answer, stops, and starts again. She relaxes gradually across the last few seconds rather than immediately.
These are short clauses and they carry most of what viewers read as acting. A model given only the start and end states will interpolate between them evenly, and even interpolation is exactly what human performance never does.
Frequently Asked Questions
How do I stop an AI actor looking at the camera?
State the eyeline target explicitly for each beat. Say she keeps her eyes on the other character throughout, or names a specific object in the room.
Can AI actors handle subtext?
They handle the behaviour that produces subtext, which is what matters. Direct the withheld action, the delayed answer, the look away, and viewers supply the interpretation.
How many performance beats fit in thirty seconds?
Two or three. More than that and the transitions become too fast to read as acting rather than as glitching.
Should I write dialogue and performance direction together?
Keep them separated in the prompt but aligned in timing, so the model can tell which is spoken and which is behaviour. Mixing them tends to produce dialogue appearing as on screen text.
Does this differ for vertical framing?
Yes, considerably. Face and eyeline dominate the vertical frame, so small performance choices carry much further and large gestures can leave the frame entirely.
What if the performance is right but inconsistent between takes?
Lock identity and wardrobe with references, then keep the performance brief identical word for word between attempts. Rewording the brief changes the performance even when the meaning is the same.
Start Directing Instead of Guessing
Every technique on this page gets easier when the take, the references and the continuity rules live in one place instead of scattered across tabs and text files. That is what Vertex is built for. Create a free account at MinionArts and start building your first microdrama scene today.
