| name | performance-direction |
| description | Provider-independent performance direction for generated video, avatar/spokesperson clips, animation, ads, film scenes, social content, and voice-led media. Use when an agent must cast or direct synthetic performers, avatars, animated characters, talking heads, AI video characters, or voice performances; plan acting beats, objective/obstacle/action, blocking, gesture, eye-line, facial expression, body language, lip-sync, voice alignment, rehearsal prompts, iteration, continuity, consent, and performance QA. |
Performance direction for generated media
Use performance direction to make a generated person or character appear to want something, react to pressure, and communicate through voice, face, gaze, and body. Do not ask a model for a generic emotion and hope it supplies acting. Give playable circumstances, beat-level actions, physical behavior, vocal behavior, continuity constraints, and a way to judge the take.
Evidence stance
Documented facts used by this skill:
- Stanislavski-derived actor training commonly analyzes a role through given circumstances, tasks/objectives, actions, beats, obstacles, and physical action; for AI use, translate these into concise, playable instructions rather than academic labels. Source cross-check: Stanislavski system overview and NYFA summary of Stanislavski questions.
- Lip-sync systems often map speech sound to visemes, not one mouth shape per phoneme. Microsoft documents visemes as visual phoneme descriptions and notes many phonemes can share one viseme; Meta's Oculus Lipsync documentation describes interpolation between visemes over time. Verified 2026-07-10: Microsoft viseme docs, Meta viseme reference.
- Professional dubbing guidance treats timing, natural language, audio level, and physical mouth/gesture sync as production constraints. Verified 2026-07-10: Netflix films/series dubbing guidelines, Netflix nonfiction simil-sync guidance.
- Current avatar tools expose different levels of performance control: some require clean training footage and consent capture, some accept a script and avatar selection, and some offer sentiment/emotional-state controls. Verified 2026-07-10: Synthesia Studio Avatar requirements, D-ID quickstart, HeyGen Digital Twin docs.
- Current AI video prompting can include character references, longer clips, extension, and batch workflows in some providers; treat capability limits as provider- and date-specific. Verified 2026-07-10: OpenAI Sora 2 prompting guide.
- Performer likeness, voice, and digital replica use raise consent, compensation, use-limit, storage, and disclosure duties. Verified 2026-07-10: SAG-AFTRA AI resource page, SAG-AFTRA Digital Replicas PDF, NAVA synthetic voice guidance.
- Provenance/disclosure can be supported by metadata standards such as C2PA Content Credentials, but provenance support varies by platform and workflow. Verified 2026-07-10: C2PA.
Empirical observations to collect on each job:
- Which performance details the chosen model obeys reliably: eye-line, gesture, expression intensity, hand accuracy, lip-sync, pauses, breath, camera blocking, and continuity.
- Whether the model responds better to actor-direction language ("tries to reassure") or visible-output language ("soft smile, shoulders lower, eyes return to lens").
- Which constraints degrade output when over-specified, especially in short clips or avatar tools with fixed pose libraries.
Production heuristics, not universal facts:
- Direct actions before emotions: "tries to win her trust" usually produces more coherent behavior than "is sad."
- Use one dominant objective per beat and one visible adjustment at the beat turn.
- Calibrate emotion on a 0-5 screen intensity scale; most business/avatar clips need 1-3, while animation, comedy, and social hooks can tolerate 3-5.
- Give every gesture a job. Random pointing, nodding, and hand waving read as synthetic.
- Lock voice timing before final avatar/lip-sync generation when the face is audio-driven.
- Repair the smallest failed layer first: script intent, voice read, facial expression, body action, camera, edit, or generation settings.
Build the performance brief
For every speaking or acting subject, write a compact brief before prompting generation:
- Medium and constraint: avatar presenter, cinematic AI video, rigged animation, generated ad, dubbing/localization, social UGC, or voice-only.
- Casting: age range, presence, vocal texture, physical energy, style reference if licensed/allowed, and what not to imitate.
- Given circumstances: where the character is, what just happened, who they believe they are speaking to, and what pressure is active.
- Objective: what they want from the viewer or other character in this beat.
- Obstacle: what makes that objective difficult.
- Action verb: what they do to overcome it, such as reassure, challenge, charm, deflect, confess, teach, invite, or warn.
- Beat turn: the moment when new information or resistance changes the tactic.
- Visible behavior: posture, gaze, hands, facial muscles or expression, breath, stillness, movement path, proximity, and touch/no-touch boundaries.
- Vocal behavior: pace, pitch range, volume, pauses, stress words, breath, laugh/hesitation, accent/dialect only when appropriate and safe.
- Continuity: where the eyes, hands, body, prop, emotional state, and voice start/end so adjacent shots match.
- Safety/rights: consent status, likeness/voice rights, permitted use, disclosure/provenance requirement, and prohibited resemblance.
Keep the brief short enough to fit into the generation prompt or handoff notes. If a provider has weak control over a dimension, move that dimension into QA and iteration rather than pretending it is guaranteed.
Direct playable beats
Beat direction should answer:
- What changed?
- What does the performer want now?
- What tactic are they trying?
- What leaks through despite control?
- What exact visible/vocal behavior should change?
Use playable action verbs instead of result adjectives:
| Weak direction | Stronger performance direction |
|---|
| "She is worried." | "She tries to sound calm while checking whether the viewer understands; her eyes hold steady, then flick down for half a beat before she recovers." |
| "Make him confident." | "He invites the viewer in as if sharing a shortcut he has tested; relaxed shoulders, direct lens contact, small smile on the first proof point." |
| "Very emotional." | "He starts controlled, swallows the next word, then lets one breath land before saying the final sentence more quietly." |
| "Friendly avatar." | "Warm onboarding host: open chest, hands quiet at waist, nod only at the user's likely concern, pace 10% slower on setup steps." |
For short generated clips, limit each shot to one or two beat turns. If the scene needs several tactical shifts, split it into multiple shots or takes.
Calibrate emotion and expression
Name the underlying emotion only after defining the action. Then specify screen intensity:
- 0: neutral task focus.
- 1: trace; readable mostly in eyes, breath, or timing.
- 2: natural professional expression.
- 3: clear dramatic/social read.
- 4: heightened genre or animated read.
- 5: caricature, panic, slapstick, or theatrical exaggeration.
Add contradiction when needed: "voice stays professional, face reveals concern." This prevents flat delivery and avoids asking the model to make every channel express the same emotion.
For faces, prefer observable cues:
- brow: relaxed, slight pinch, lifted inner brow, skeptical brow drop;
- eyes: direct lens, tracking another character, soft focus, brief eye dart, slower blink;
- mouth: closed-mouth smile, smile suppressed, lips press, jaw releases, half-laugh;
- head: still, small nod, tilt, recoil, lean-in, turn away.
Avoid stacking too many expression cues. A face cannot convincingly do "wide smile, clenched jaw, whispered grief, aggressive stare, relaxed warmth" in one beat unless the contradiction is intentional and physically coherent.
Blocking, gesture, and eye-line
Blocking is performance meaning in space. State where the performer is, how they orient to the viewer/partner/camera, and why they move.
Use these defaults:
- Direct-to-camera presenter: lens is the listener. Use steady eye contact, occasional thinking glances away, and deliberate re-entry to lens for key claims.
- Dialogue: define each character's eye-line target and whether they avoid, challenge, seek, or break contact.
- Product/social ad: let the body discover or demonstrate the product; hands enter only when the story needs proof, comparison, or invitation.
- Film scene: block movement around power, avoidance, attraction, secrecy, or discovery. Distance and stillness can be more expressive than constant movement.
- Animation: make silhouettes and pose changes readable; use anticipation before large action and settle after emotional changes.
Gesture rules:
- Give each gesture a verb: reveal, stop, measure, protect, count, invite, dismiss.
- Match gesture scale to frame size. Close-ups need micro-gestures; wide shots can use full-body action.
- Do not ask for constant hand motion in avatar tools unless the tool supports gesture control and the brand style tolerates it.
- Keep hands away from faces, logos, small props, and occluded regions when the model is prone to artifacts.
Eye-line continuity matters. For multi-shot scenes, record who/what each character is looking at at the start and end of every shot. In AI video, inconsistent eye-line is often a continuity failure, not just a face failure.
Align voice, lip-sync, and face
Voice performance drives perceived acting in avatar and voice-led media. Direct:
- intention: what the speaker is doing to the listener;
- pace: steady, urgent, deliberate, conversational, compressed;
- pauses: where the thought changes, not just where punctuation appears;
- emphasis: one or two key words per sentence;
- breath: calm inhale, held breath, laugh-breath, recovery breath;
- contour: where energy rises, falls, or resolves;
- articulation: crisp, relaxed, mumbled, hesitant, but only if it serves clarity.
For synthetic voices, write performance markup in the syntax the provider supports; if syntax is unknown, use plain text notes adjacent to the line. Verified 2026-07-10, ElevenLabs documents pacing as a major part of voice design and advises clear pacing language for voice outputs: ElevenLabs Voice Design.
For lip-sync/avatar generation:
- Finalize or nearly finalize the script before generating a face.
- Generate or select the voice take.
- Check word timing, pauses, and emotional arc.
- Generate avatar/lip-sync from the selected audio or script.
- QA mouth open/close starts, final closures, bilabials (p/b/m), fricatives (f/v), and visible jaw rhythm.
- If mismatch appears, repair the audio timing or line length before over-directing the face.
For dubbing/localization, maintain meaning and natural speech while respecting original timing and visible mouth/gesture openings. If perfect internal mouth-shape match makes language unnatural, choose the sync type explicitly: lip-sync, simil-sync, voice-over, narration, or subtitle-supported dub.
Avatar and animation constraints
Do not direct every medium like live action.
Avatar presenter:
- Strengths: consistent framing, direct address, scalable delivery, training/explainer/social presenter formats.
- Risks: repetitive gestures, fixed posture, uncanny eye contact, emotional flattening, mismatch between voice clone and face.
- Direction style: concise presenter intent, frame, pace, sentiment, gesture density, allowed camera angle, and brand energy.
Talking photo:
- Strengths: fast personalization and face-led address.
- Risks: limited body language, weak gesture, head/eye artifacts, identity sensitivity.
- Direction style: voice-first; keep expressions simple and avoid complex physical action.
AI cinematic character:
- Strengths: richer movement, environment, camera, and scene dynamics when supported by the model.
- Risks: continuity drift, inconsistent acting across shots, uncontrolled hands/eyes, over-literal emotion.
- Direction style: shot-by-shot beats, physical blocking, camera relation, start/end continuity, references where permitted.
Rigged 2D/3D animation:
- Strengths: precise pose, gesture, timing, repeatability, exaggeration.
- Risks: stale poses, floaty transitions, weak weight/anticipation, lip-sync mismatches.
- Direction style: pose keys, timing, ease, silhouette, anticipations, holds, settles, facial control names if available.
Voice-only:
- Strengths: maximum control through rewrite, casting, pacing, and edits.
- Risks: monotonous read, overacting, unclear speaker relationship.
- Direction style: line-by-line intention, listener image, tempo map, pause map, emphasis, pickups.
Casting, likeness, and consent
Before creating or directing a synthetic human, decide whether the performer is:
- fictional and non-identifiable;
- a stock/licensed avatar;
- a custom avatar based on a consenting person;
- a known public figure, employee, customer, influencer, actor, or deceased person;
- an imitation of a real voice, face, style, or performance.
Require written scope for custom or recognizable likeness/voice work:
- who gave consent and whether they are authorized to grant it;
- what media, script type, territory, duration, platform, and edits are allowed;
- whether the voice/likeness can train models, generate new lines, or be reused later;
- compensation, revocation, storage, exclusivity, and deletion terms when applicable;
- disclosure/provenance plan for synthetic media.
If consent is missing or ambiguous, redirect to a fictional/non-identifiable performer, licensed stock avatar, or human-recorded performance. Do not create a "soundalike" or "lookalike" workaround for a real person when the intent is recognizability.
Rehearse through prompts
Use prompt rehearsal before expensive final renders:
- Table read: generate or write two voice takes with different actions, not different emotions.
- Pose rehearsal: ask for still frames or key poses when available; check silhouette, gaze, hand placement, and expression intensity.
- Beat rehearsal: generate a short version of the scene with only the beat turn.
- Continuity rehearsal: test adjacent shots with the same start/end states and references.
- Final performance: add production detail only after the acting read works.
Iteration notes should be specific:
- "Reduce smile from 3 to 1; keep warmth in eyes, not teeth."
- "Hold eye contact through the objection, then look down only after the denial."
- "Gesture once on the proof point; hands still for the emotional admission."
- "The line is too fast for the mouth closure; add a 200-400 ms pause before the final word or choose a slower voice take."
- "The character is displaying fear, but the objective is to persuade. Re-direct as 'tries to reassure while managing fear.'"
Performance QA
Review the output with the sound on, sound off, and audio-only.
Sound on:
- Does the performer pursue a clear objective?
- Does the voice match face/body energy?
- Are beat turns audible and visible?
- Do pauses feel like thought, not latency?
- Are dubbed/avatar mouth starts and endings acceptable for the chosen sync standard?
Sound off:
- Can you infer relationship, attention target, and pressure?
- Are gaze, posture, and hands motivated?
- Does the expression intensity match genre and brand?
- Are there uncanny loops, frozen blinks, hand artifacts, or random nods?
Audio-only:
- Does the read have a listener, tactic, and arc?
- Are key words stressed without sounding salesy or theatrical?
- Is the pace appropriate for platform, information density, and captions?
Continuity:
- Start/end emotion, body orientation, eye-line, prop state, wardrobe, and voice tone match adjacent shots.
- Character references, avatars, and voices are identified in the handoff.
- Consent/provenance notes travel with the asset.
Handoff format
For each performer or character, hand off:
Performer: <name or asset id>
Consent/rights: <fictional/licensed/custom consent status; allowed uses; disclosure>
Casting read: <age/presence/vocal/body energy; prohibited resemblance>
Scene objective: <what they want>
Obstacle: <what resists them>
Relationship/listener: <who they believe they address>
Beat map:
00:00-00:03 <action, expression intensity, gaze/body/voice>
00:03-00:07 <beat turn and new tactic>
Voice notes: <pace, pauses, emphasis, breath, pronunciation>
Blocking/gesture: <position, movement, hands, props, eye-line>
Continuity in/out: <start and end states>
QA priorities: <highest-risk performance details>
Example: avatar onboarding clip
Production intent: 25-second SaaS onboarding video for a warm but credible avatar presenter. Medium: script-driven avatar with direct-to-camera framing.
Performance brief:
Performer: licensed stock avatar, mid-30s, calm technical host.
Objective: reassure a new user that setup is quick and safe.
Obstacle: the user expects integration setup to be annoying.
Action: demystify, then invite.
Emotion scale: warmth 2, urgency 1, confidence 2.
Eye-line: lens as user.
Gesture density: low; one small open-hand gesture on "two minutes."
Voice: conversational, 145-155 wpm, slight smile in tone, pause after first sentence.
Consent/rights: stock avatar within platform license; disclose synthetic presenter if required by channel policy.
Script/performance direction:
[steady lens contact; relaxed shoulders]
"You do not need to rebuild your workflow."
[small pause; softer smile, warmth 2]
"Connect your calendar, choose the approval rule, and we will show you the first safe draft before anything goes live."
[one open-hand gesture on "first safe draft"; pace slows 8%]
"Most teams finish setup in about two minutes."
[return hands still; invite, not pressure]
"Let us do the first pass together."
Why it is structured this way: the avatar gets one objective, one obstacle, a restrained gesture plan, and voice timing that supports lip-sync and trust. Likely failures: over-smiling, random nodding, too-fast line two. Repair by reducing expression intensity, removing extra gesture requests, or splitting the long sentence.
Example: generated cinematic two-character scene
Production intent: 8-second dramatic shot for a short film teaser. Medium: AI video, over-the-shoulder close-up, no celebrity likeness.
Prompt-ready direction:
Shot: close over Mira's shoulder toward Jonas at a rain-streaked bus shelter at night, 50mm, shallow depth of field. Fictional characters, no resemblance to real actors.
Given circumstances: Jonas has just been caught hiding the train ticket; Mira is waiting for an explanation off camera.
Jonas objective: persuade Mira not to leave.
Obstacle: he knows his excuse is weak and she has stopped trusting him.
Action: confess just enough to keep her there.
Beat 1 (00:00-00:03): Jonas holds Mira's eye-line slightly left of lens, jaw tight, breath held, face controlled at regret 2.
Beat turn (00:03): he sees she is about to walk away.
Beat 2 (00:03-00:08): his shoulders drop, eyes flick to the ticket then back to Mira; he says quietly, "I bought two." No smile. Voice low, slower on "two."
Blocking: Jonas remains still except the ticket hand lowers into frame; Mira is a blurred shoulder foreground, no speaking.
Continuity out: Jonas looking at Mira, ticket visible at chest height, rain and neon consistent.
QA: eye-line to Mira, hand artifact on ticket, no melodramatic crying, mouth closes after "two."
Why it is structured this way: the scene is directed through objective and obstacle, not "sad man." It gives the model one physical reveal and one emotional shift. Likely failures: ticket hand distortion, eyes to camera, overacted tears. Repair with a closer crop, clearer prop silhouette, or reduced emotion scale.
Example: animated explainer character with voice
Production intent: 15-second cartoon character explains why password managers help. Medium: rigged 2D character with TTS and simple viseme lip-sync.
Direction:
Character: small raccoon office helper, curious and practical, no real-person likeness.
Objective: convince the viewer to stop reusing passwords.
Obstacle: the viewer thinks password managers are extra work.
Action: make the danger concrete, then offer relief.
Animation constraints: use three poses only: lean-in warning, count-on-fingers, relaxed thumbs-up. Keep paws below chin to avoid mouth occlusion.
Emotion scale: concern 3 for first beat, relief 3 for final beat; animation style allows exaggeration.
Voice: bright, quick but clear; pause after "same key."
Line map:
00:00-00:05 lean-in warning, eyes wide, paws still:
"Using one password everywhere is like using the same key for your house, car, and diary."
00:05-00:10 count-on-fingers, one paw gesture only:
"A password manager makes different keys and remembers them for you."
00:10-00:15 relaxed thumbs-up, smile 3, slower final phrase:
"You remember one master password. That is the deal."
Lip-sync QA: check open vowels in "everywhere" and final mouth closure on "deal"; if weak, slow the final sentence or regenerate visemes from final audio.
Why it is structured this way: the acting is simple enough for a rig, gestures do not fight lip-sync, and each pose has a persuasive job.
