@aswless_854771076/ai_short_studio_cli 0.1.97 → 0.1.99

package/skills/using-vvicat-ai-short-studio-cli/references/hypit/playbooks/craft/video-direction.md

@@ -0,0 +1,433 @@
+# Directing generated video
+
+Read this before writing or adapting a video prompt, performance Recipe or action. An image can
+establish who and where; generated video brings the Script and visual idea to life through speech,
+interaction, camera behavior or silent action.
+
+For creator-led social video, favor an appealing, engaged performance: the speaker responds to what
+they are saying and draws someone into the thought. Voice, expression and movement carry that
+relationship together. Expressiveness concerns how clearly the attitude comes through; it can be
+outgoing or restrained.
+
+## Let footage carry temporal behavior that matters
+
+Choose generation inputs from what the new shot must actually realize. A still can establish a
+person, object, place or composition; prose can direct an action and its expressive purpose. When a
+particular unfolding in time is essential to the result—its performed rhythm, path, contact,
+coordination or camera relationship—and prose and stills would lose that information, use the
+relevant source footage as a video reference to a model that accepts it. This judgment follows the
+action the work needs, whether or not the user names a generation method. A broadly described action
+can leave more freedom to the model when its exact movement is not the point.
+
+Replacing a person is a common case; the target may instead change an object, world, words or sound.
+A comic speaking performance can depend on the relation between vocal phrasing, face, gesture and
+camera angle; dance on coordination, a fight on contact, and a fixed-camera mishap on how its cause
+and payoff develop within the view. Watch and listen to learn what the source contributes. Its vocal
+manner can inform direction, while Script and [voice direction](voice-direction.md) establish the
+target's words and voice. The video reference can carry temporal behavior while image references
+establish the new visible facts and the prompt directs their relationship.
+[Transformation](../../creation/transformations.md#reshape-an-event-from-footage) owns what the
+new work keeps and reshapes. Watching a reference to understand a work is distinct from connecting
+a selected excerpt as an actual generation input.
+
+Choose an excerpt that contains the useful action and, when relevant, its preparation or settling.
+Its range follows the relationship to preserve, the intended new duration and the selected model's
+reference limits. Connect the actual footage in Source; describing it in prose does not supply
+its temporal behavior to the model. The selected model and Provider determine supported combinations
+and media requirements; read their installed vocabulary and package documentation for the exact
+request. Seedance requires `person-reference="true|false"` on each image/video reference: true if
+the supplied material contains a person, false otherwise.
+Inspect the selected excerpt, not just its first frame. Audio must omit the field; the installed
+Seedance README owns the corresponding required first/last-frame attributes.
+
+For a reference-led recreation, a depth prepass is mandatory before final shot generation. Generate and
+inspect the depth video first, then pass it as the motion/structure/camera reference; an optional selected
+character board may additionally carry identity and appearance. Never omit this edge from the request
+graph; choose another model or Endpoint when the current route cannot consume it.
+
+Before compiling a recreation request, approve the selected character, scene and prop assets and the
+shot's depth video, then bind their actual files with the video prompt in one ordered per-shot manifest.
+Add audio and a reviewed storyboard when the shot plan requires them; ordinary recreation shots do not
+need two boards.
+[Asset and shot preproduction](asset-and-shot-preproduction.md)
+owns this gate and the different responsibility of each reference. A complete asset library does not
+mean every asset belongs in every request; use the views that know this shot without silently dropping a
+required identity, world, prop-state, planned storyboard or depth role.
+
+Lock a recreation's visual baseline before generating shots: source exposure and white balance, skin
+texture and diffusion, contrast and saturation, lens/field of view, depth of field, camera height and
+movement character. Carry the same concrete baseline into every related request. Do not pass original
+keyframes or previous generated tails as routine recreation references. Preserve identity, costume,
+body proportions, screen direction, held-prop state,
+key-light direction and the source shot's end pose across that handoff.
+
+Direct physical actions through contact states rather than action names alone: anticipation, limb and
+joint path, contact point, transferred weight, prop response, recovery and a briefly stable tail. Split
+an interaction when one request would need several independent contacts or exact hand choreography.
+
+## Build natural motion from causes, not adjectives
+
+Do not use `natural movement`, `realistic motion` or `cinematic motion` as a substitute for direction.
+Give every visible person one current primary action, an attention target, a trigger for reacting and a
+clear state while uninvolved. In a group, name people by stable position or identity and stagger onset,
+amplitude, gaze and reaction latency unless the Script truly calls for marching, ritual unison or a
+shared startle. A group that turns, nods, blinks or starts walking on the same frame reads as generated.
+
+Stillness is not a frozen frame. At a shot size where the behavior is visible, let breathing move the
+chest and shoulders slightly, let gaze refocus through small eye movements, allow irregular independent
+blinks and carry expression through a continuous transition. Use only a few of these cues and keep
+background behavior below the narrative subject; a list of periodic micro-actions creates another kind
+of mechanical performance. Voice-over keeps visible lips closed, but the face, eyes, breath and posture
+remain alive.
+
+Give larger movement a physical chain: planted support, center-of-mass transfer, hip or shoulder lead,
+joint path, acceleration, deceleration, follow-through and recovery. After contact, keep the correct
+grip, load and connected-object response. Hair, hems, loose sleeves, accessories and carried objects are
+secondary motion: they lag the driving body or wind, travel at smaller amplitude and settle with damping.
+Do not add windless flutter, whole-body phase locking or an instant stop. The setting may respond only to
+an evidenced force—a cushion compresses after sitting, a door creates a brief draft, a footfall disturbs
+nearby dust—not because the shot needs generic movement.
+
+Keep the prompt's motion budget small: one narrative primary action and only the secondary motions that
+make its cause, weight or reaction legible. Split a shot when several exact interactions compete. Give
+the camera one motivated curve with a start, speed development and end; a locked-off camera never drifts.
+This follows the practical separation of subject, camera and scene motion in Runway's official Gen-4
+guidance, the structured subject/action/context approach in Google's Veo guidance, and the prompt-
+extension practice published with Wan. These sources support richer, explicit motion description; they
+do not make text alone a substitute for depth, motion tracks or other temporal references when exact
+timing matters.
+
+Primary references:
+
+- Runway, `Gen-4 Video Prompting Guide`: https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide
+- Google DeepMind, `How to create effective prompts with Veo 3`: https://deepmind.google/models/veo/prompt-guide/
+- Wan-Video, `Wan2.1`: https://github.com/Wan-Video/Wan2.1
+- Lightricks, `ComfyUI-LTXVideo` motion-track workflows: https://github.com/Lightricks/ComfyUI-LTXVideo
+
+## Turn subtext into gaze and facial behavior
+
+An internal state is not a visible instruction. Convert it into a causal performance chain: the person
+receives a word, sound, sight or touch; attention moves; they try to preserve a social mask; one small
+feature leaks the change; they decide or act; a residual expression survives into the cut. Stop when the
+Script, storyboard or approved character profile supplies no evidence for the next link. Do not invent a
+secret motive merely to make the shot feel dramatic.
+
+Give gaze an object and screen geography. Name the person, prop or off-screen position being watched;
+state whether the character holds, glances, averts or returns; and choose whether the eyes lead the head
+or the head turns with them. Preserve the off-screen target across the axis. Looking into the lens is a
+specific direct-address or POV choice, not a default sign of intensity. Direct the listener too: receiving,
+judging, resisting or anticipating can be legible through attention and timing without copying the
+speaker's movement.
+
+Facial direction should describe visible change instead of a broad emotion label. In a close shot, use
+one or two readable regions—brow tension, eyelid narrowing, a held lower lip, one lip corner losing its
+smile, nostril tension or a jaw set—and give the change a low-amplitude onset, apex and release. Do not put
+FACS Action Unit numbers into a model prompt; FACS is useful here because it separates observable muscle
+components, not because its codes are model vocabulary. Do not stack the whole face or hold a
+micro-expression as a pose. At medium distance, carry the same beat through gaze, head, breath and jaw;
+at wide distance, use posture, pause, interpersonal distance and blocking.
+
+Bind these choices to the character profile. `eyeline_behavior`, `blink_rhythm`, `stress_response`,
+triggered `habitual_actions` and `forbidden_performance` travel with the selected character reference as
+`performance_constraints`. A habitual action appears only when its recorded trigger occurs, while a
+forbidden performance remains forbidden in every shot. Profile behavior supplies continuity; the current
+dramatic stimulus determines whether and how it becomes visible.
+
+OpenFace's public implementation treats gaze, head pose and facial Action Units as distinct measurable
+channels, while FACS decomposes facial movement into anatomical components. GoHD likewise separates pose,
+gaze and expression driving. These are useful design boundaries: coordinate the channels around one
+dramatic beat, but do not collapse them into one synchronized facial gesture.
+
+Primary references:
+
+- Paul Ekman Group, `Facial Action Coding System`: https://www.paulekman.com/facial-action-coding-system/
+- OpenFace, `Action Units` and gaze APIs: https://github.com/TadasBaltrusaitis/OpenFace/wiki/Action-Units
+- GoHD official implementation: https://github.com/Jia1018/GoHD
+- EMOCA official implementation: https://github.com/radekd91/emoca
+
+For a longer action, distinguish the duration of source footage the model can accept from the
+duration it can produce in one request. Use one request when the model's capability and the work's
+continuity allow it. A forty-second fight might contain several exchanges and reversals, or its
+uninterrupted choreography might be the point. If separate requests serve the work, select excerpts
+around meaningful action or camera relationships and carry forward the target
+people, objects and world through useful references. A literal boundary frame helps when that exact
+continuity is intended; it is not needed for every new passage. The accepted outputs establish the
+target's actual time. When uninterrupted continuity is essential, use a capability that can carry
+it or explain what an edited interpretation changes. When an accepted wordless action carries the
+passage, [empty Segments](../../production/media.md#empty-segments-use-their-media-boundaries)
+let that material establish its time without word tokens.
+
+[Reference relationships](generated-dependencies.md) owns how images carry identity and world
+forward across requests. This page owns the decision to give a video request temporal evidence.
+
+## Direct the reason for an action
+
+Give the passage a clear expressive intention, then specify the few details that decide how it
+lands. Start with how the speaker regards the subject and relates to the listener: eagerly sharing
+a discovery with a friend, weighing an ambitious claim with skeptical interest, or affectionately
+roasting something they know well. Such directions suggest voice, expression, posture and rhythm
+together. Their value is the performable relationship they contain.
+
+As with [Image direction](image-direction.md#two-ideas-behind-every-sentence), high-level
+language should carry sensory direction. The model can realize "admiring but incredulous" across a
+whole passage. Words such as "lively" or "expressive" set an energy level; the speaker's particular
+response to the meaning gives that energy direction. Locate the consequential thought in the Script
+and make its expression concrete: what earns an accent, changes the face, or moves the body?
+Voice, expression and gesture can reinforce one response, with whichever channels make it legible.
+A single revealing reaction can carry the thought; specifying all channels at every phrase adds no
+inherent value.
+
+For example, alongside a Script that questions a price and ends with an ironic compliment:
+
+```text
+She is venting to a friend with affectionate exasperation at how overpriced this is. At the price,
+her eyebrows rise and her open palm turns upward with an incredulous vocal accent. Let the final
+compliment land with a teasing smile in her voice.
+```
+
+The first sentence directs the entire encounter. The details make the price and the ironic turn
+read clearly; the remaining phrasing and movement can grow naturally from that attitude. A detail
+earns its place when it clarifies a judgment, reveals a response or preserves an important physical
+relationship. Attaching a hand movement to every phrase adds choreography without necessarily
+strengthening the expression.
+
+Let the thought determine whether the attitude develops or holds: an example may win a skeptical
+speaker over, while a firm argument may sustain the same conviction throughout. When adapting a
+reference, recover what prompts the reaction and how it lands, then find the corresponding thought
+in the new Script. This preserves the expressive relationship while letting the new words call for
+their own movements and phrasing. Phrase-level emphasis should fit the performed language's tones
+and cadence; a source-language stress pattern is not a universal performance instruction.
+
+Carry the character, voice and useful physical relationships across Takes while directing the
+attitude each passage calls for. The same speaker can invite, question, tease and persuade as the
+argument develops. Shared direction can preserve their manner; the passage supplies the particular
+response. Carry forward production continuity, and reconsider the expressive intention whenever the
+thought changes. Stable framing leaves room for changes in face, voice and posture.
+
+## Give restraint an expressive purpose
+
+Seriousness can mean weighing a claim, insisting on a fact or challenging an assumption. A relaxed
+person can tease, dismiss or show interest with very little effort. Choose the behavior that conveys
+that attitude. A small, deliberate response can carry more conviction than constant movement.
+
+For a serious judgment:
+
+```text
+She weighs the claim with skeptical interest. Her brow tightens at the price; she gives the number
+a pointed emphasis, then lands her judgment with a small, decisive nod.
+```
+
+For a languid response:
+
+```text
+He lounges back in the chair, amused that anyone finds this impressive. He draws out the setup
+with a sideways glance, then tosses off the verdict in a light, dismissive tone.
+```
+
+These are different intentions, not prescribed poses for every serious or relaxed person. Breathing,
+blinking and small posture adjustments supply ordinary life; the speaker's response to the content
+supplies the performance. Direct that response even when the body stays almost still.
+
+Contrast describes a relationship between things; it leaves the performer's attitude unspecified.
+Turn the chosen relationship into something the person can express: "She tries to sound composed,
+but cannot quite hide her pride in the result" gives voice and expression a shared direction.
+Humor reaches the viewer through the words, situation, response or delivery. A straight-faced
+performance can sharpen a clear joke through pointed emphasis or a dry turn of phrase; an unusual
+appearance alone does not make an even explanation funny. [Treatment](../../creation/brief.md#treatment-is-the-directors-answer)
+owns how the premise becomes a viewer experience; action directs the person's part in creating it.
+
+## Ground actions in the generated scene
+
+The prompt describes what this generation should make visible over time. Give a performer physical
+relationships with the camera, people, props and parts of the setting that actually exist in the
+generated scene. Translate a later graphic's framing needs into the camera view, gaze or gesture
+the video should perform. Give the later Caption, icon or product card its own content and events
+through composition. A prop or display intended to exist within the generated scene can be described
+as that actual object and connected through the relevant references.
+
+The spoken subject and the visible action have separate responsibilities. A speaker can discuss
+objects, quantities, transformations or an imaginary demonstration while the camera records only
+their performance in the referenced setting. Preserve the words in Dialogue; use action to describe
+the person's response to those words. A quoted phrase can locate a vocal accent or a hand beat
+without asking for its meaning to appear as an object, written label or event in the shot.
+
+Make the performer the subject of the direction. "She gives the promise a satisfied accent and a
+small nod" identifies what the camera and microphone should capture. "Make the idea tangible" or
+"show that it works" leaves the means of demonstration open. Complete that translation into the
+intended voice, expression or physical interaction before writing the request. Read the assembled
+prompt as one account of the generated scene; a general no-text instruction does not resolve an
+ambiguous invitation to demonstrate the spoken content visually.
+
+Capable video models can turn figurative wording into literal objects, events or transformations. Use
+concrete visible language for intended gaze, gesture, movement, camera behavior and cuts when a
+metaphor would introduce the wrong scene content. Social attitude and aesthetic shorthand remain
+useful when they direct performance; an imagined object or event belongs in the prompt when it should
+truly appear in the generated world.
+
+Give the performer an attitude they can express, such as questioning a claim with skeptical interest.
+The broader purpose of persuading an audience belongs in Treatment; translate it into this person's
+delivery and interaction. Decide whether a described action really happens or whether a gesture is
+intended, then state that choice. [Direction and its inputs](../../production/system.md#give-each-part-the-direction-it-can-realize)
+connects those performance choices to the whole composition.
+
+Natural emphatic gestures are usually more reliable than asking fingers to display an exact number.
+Let speech, Caption or MG convey the quantity while pointing and hand actions serve the performance.
+An encounter also has edges: someone interrupted can first be occupied, and someone finishing can
+begin to leave. Small causes make a clip feel like a piece of life rather than a pose bounded by the
+encoder.
+
+## Prepare footage for subject isolation
+
+When the work needs a moving silhouette, choose a background-removal method that can process the
+actual footage. For new generation, a continuous, evenly lit chroma backdrop can support a chosen
+keying workflow. Choose a color separated from the person's hair, clothing and carried objects;
+direct the visible body extent, performance and stable backdrop as facts of this recording. This
+is a material-preparation choice for that use. Existing footage can instead use a suitable matting
+capability, and an opaque camera view can remain useful through cropping and reframing.
+
+The moving result needs its own alpha-producing operation. A transparent reference still alone
+does not establish transparency in generated video. Keep the original when the final work also uses
+its setting. [Media preparation](../../production/media.md#keep-original-and-processed-material-explicit)
+owns processing, alpha-preserving normalization and explicit use of those outputs. Judge the moving
+edge against its intended background, including shoulders, hair, hands, spill and changes in opacity.
+
+## Give each input its own responsibility
+
+| Input | Responsibility |
+| --- | --- |
+| Character-and-scene references | appearance, setting, framing and the physical state to preserve |
+| Product, interface or other factual references | the visible facts that need continuity or exactness |
+| Source-footage references | performed, physical or camera relationships whose timing should guide the new shot |
+| Recurring voice references | a speaker's intended voice identity when the model accepts them |
+| Script dialogue | the exact words, intended pronunciation and speaking turns for a visible performance |
+| Prompt Kit or Recipe | a reusable prompt relationship that fits this kind of work |
+| Passage direction | attitude, vocal delivery, attention, physical interaction, camera behavior and motivated cuts |
+
+Keep those responsibilities explicit in Source. Prompt prose does not create a media edge, and a
+reference does not explain which fact it should preserve. A selected model may accept only some of
+these inputs; use its installed vocabulary and package-local documentation for the exact request.
+Generated source footage must not contain incidental subtitles, captions, titles, labels, logos,
+watermarks, UI text or other unrequested readable text. Approved story text that physically belongs in
+the shot, including a countdown, may be generated when the Brief calls for it.
+
+For each exact countdown, clock, message or difficult changing state, make a temporal storyboard before
+video generation. Default to six cells, while selecting another count when needed for accurate state
+coverage. Give each reading or required state its own cell, approve the original-resolution pixels,
+then attach the board as a semantic reference and direct the exact sequence and timing. The output remains
+one normal full-frame video and must not reveal the grid. Reject malformed output and regenerate after
+improving the prompt or board; never repair generated video with painting, inpainting, compositing,
+source-footage patches or local region replacement.
+
+For a visible A-roll performance, the generated video normally carries the person's picture, exact
+Script delivery and sound together. In recreation mode, give the request the Segment's dialogue and
+an audio reference when the reviewed shot plan requires one; general non-recreation modes may use a
+recurring Voice Reference when supported.
+[Voice direction](voice-direction.md) owns the casting and sample that establish who the person sounds
+like. The passage's direction gives that voice its current attitude and delivery.
+[Script pronunciation](../../creation/script-and-time.md#write-the-intended-pronunciation) explains
+how names and abbreviations receive the intended reading while keeping their display spelling.
+Independent speech is a different A-roll construction, described in
+[Voice and performance](voice-and-performance.md).
+
+Final generated shots default to joint picture-and-audio generation. Wordless action still carries the
+directed ambience and physical sound unless the Brief explicitly requires silence or the Endpoint lacks
+native audio; record that waiver before generation. For visible speech, measure the dialogue first, size the
+request around that performed duration, include the exact Script line and direct voice, mouth movement,
+breath and scene sound in the same request. Seedance defaults `generate-audio` to `true`; for H3, populate
+its audiovisual prompt sections and attach a reviewed audio reference when the recreation shot requires one. When speech must be produced
+separately, plan explicit lip synchronization and review the actual phoneme-to-mouth result before the
+Take is accepted.
+
+When the source alternates languages, split analysis by utterance or speaking interval and confirm each
+turn's language independently. Write the exact original line and language code into the joint audiovisual
+request. Do not use the file's dominant language for every turn, infer speech from translated subtitles,
+or let the model translate or switch the declared language.
+
+For silent B-roll, direct the visual event and omit speaking identity that the shot does not use. A
+listener or reaction shot can remain silent while still breathing, noticing, adjusting posture or
+responding through expression. [B-roll](b-roll.md) owns its editorial relationship to the underlying
+performance.
+
+## Choose the generation relationship
+
+Seedance 2 Mini at 720p is the usual starting point for generated performances, balancing capability
+and cost. Choose for the intended shot and the user's available services; the selected model's
+installed vocabulary owns supported resolutions, references and request lengths.
+
+Generated-video models commonly expose some combination of three relationships:
+
+| Relationship | When it fits |
+| --- | --- |
+| Text-directed video | The intended world and action do not depend on an existing visual identity, composition or motion reference. |
+| First-/last-frame video | The shot must begin or end at a particular authored image. |
+| Reference-directed video | Images, video or audio should guide identity, world, voice, motion or camera language without declaring literal endpoints. |
+
+Reference-directed generation is the strong starting point for controlled Hypit production because
+useful images can already establish people, scenes, products, composition and visual continuity. The
+requested performance then makes that world move. Text-directed or first-/last-frame generation remains
+useful when its relationship is genuinely the one the shot needs; a last frame belongs when arriving
+at that exact image is part of the intended action.
+
+Model-specific Surfaces and Prompt Kits are implementations of these relationships, not the Craft
+itself. Kits produce ordinary Text; Source connects that Text and the actual references to the selected
+model request. The current Distribution may provide model-specific Kit packages such as
+`@hypit/seedance-kits`; their package documentation owns exact imports, slots, Recipe choices and
+reference order. [Prompt Kits](../../production/prompt-kits.md) explains using or authoring a Kit
+without making it a model wrapper.
+
+Read the selected template's wording when choosing its Recipe. Composition, camera, edit rhythm,
+performance and gesture choices shape different aspects of the footage. Choose them to support the
+intended delivery, and use action Text for the passage's particular meaning and reactions.
+
+## Direct camera and cuts as part of the passage
+
+For ordinary direct-to-camera social video, prefer pause-trim jump cuts at phrase boundaries. This
+edited rhythm keeps the explanation moving and gives the footage the immediacy of a creator's own
+cut-down recording. Stable framing, expressive acting and frequent edits are compatible choices.
+Use the selected Kit's documented Recipe choice to express that rhythm.
+
+A podcast can cut with a speaker or toward a meaningful reaction. A street interview can favor the
+guest and use the interviewer view for a reaction. An unfolding action or a held reaction may gain
+its effect from a continuous shot; choose that temporal shape when it serves the passage.
+
+One generated video can contain multiple shots, several speaking turns or a split-screen composition.
+One Segment is not one speaker turn or one camera shot. Conversely, several Takes can reuse the same
+character-and-scene image and meet at natural editorial cuts. A genuinely continuous shot calls for
+the model relationship and direction that preserve that action. See
+[Reference relationships](generated-dependencies.md) when deciding which visual or motion evidence
+should condition each request.
+
+When supported by the selected model, several camera images can guide one request through different
+views or locations. Direct which people, settings and states each image contributes and how the
+passage develops among them. Reference images in this relationship guide the generated footage;
+they become literal first or last frames only when that is the chosen generation relationship.
+Choose the request's span from the intended performance and the model's supported duration and
+reference inputs, not from a one-image-per-shot correspondence.
+
+A prompt-directed jump cut asks the generator for an edited rhythm; it does not inspect or trim the
+returned media. When produced footage needs a deterministic cut, speed change or trim, use the
+corresponding media operation and align the edited result before it becomes a SemanticTake.
+
+## Size the request around the delivery
+
+Use `hypit measure` on a spoken Segment at its intended pace, including time for meaningful
+interaction, pauses and actions. [Script and time](../../creation/script-and-time.md#measure-before-choosing-durations)
+owns the command, rounding and the relationship between estimated duration and real aligned time.
+Choose that pace from the intended performance and carry it into the voice and passage direction.
+For brisk social delivery with trim cuts, `fast` is a useful starting choice; the name `normal`
+does not make it the right rhythm for every piece. A shorter duration gives the words less room,
+while the direction still supplies the stresses, attitude and reactions that make them engaging.
+Measurement sizes the words; emphasis, attitude and motivated reactions give their delivery character.
+The target's delivery determines how much generated media the passage needs; the reference video's
+seconds help explain its rhythm without becoming the target duration automatically.
+
+Read the selected video model's supported request lengths from its installed vocabulary. Those values
+constrain one generation request without defining the finished work's Segment structure or edited
+beats. When the estimate does not fit one request cleanly, reconsider the performable passage: a brief
+question and answer may belong together, a short line may gain a meaningful reaction, and a long
+exchange may divide where its thought turns. Preserve the intended meaning and energy, then choose a
+supported duration.
+
+After production, normalized media supplies the real envelope and alignment supplies word positions.
+Use that material and timing to compose the piece. [Composition review](../../production/review.md)
+owns judging how Caption, MG, B-roll and other layers work with the produced performance.