Communitygithub.com

performance-direction

Preserve the expressive actions, reactions, and timing that make a reference video work; design original visual gags or hooks; and carry those choices into shots, prompts, edits, and review. Use for performance adaptation, lost facial or body detail, weak visual highlights, and performance-focused production review.

performance-direction란 무엇인가요?

performance-direction is a Claude Code agent skill that preserve the expressive actions, reactions, and timing that make a reference video work; design original visual gags or hooks; and carry those choices into shots, prompts, edits, and review. Use for performance adaptation, lost facial or body detail, weak visual highlights, and performance-focused production review.

지원 대상~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/SharingYu/video-performance-skills/tree/HEAD/skills/performance-direction

즐겨 사용하는 AI에게 물어보기

이 에이전트 스킬이 미리 로드된 새 채팅을 엽니다.

문서

Performance Direction

Make performance observable: what attracts a character's attention, how they respond, what changes the audience's understanding, and how the moment lands. Completing the plot action or matching two endpoint images does not establish that the performance survived.

For a local video, use the host's available observation tools before making reference claims. The optional video-observation companion helps cover shots and revisit uncertain interactions; it is not required for supplied text or images. Match the user's requested granularity: broad attention, body language and reciprocal action are often enough. Do not turn that request into facial-muscle measurement.

Respond in the user's language. Preserve their scope, characters, style, and decisions. Use the smallest relevant mode:

RequestReadDeliver
Adapt a reference's acting or visual appealReference performanceEvidence-based event cards, protected details, and allowed changes
Develop or repair an original visual highlightOriginal highlightsA mechanism or useful alternatives, a reasoned choice, and its event card
Translate or review an approved idea through productionProduction handoffTraceable shot/prompt/edit requirements or scoped review findings

Do not expand a local repair into a new story. Mark missing evidence as unknown; ask only for information that materially blocks the requested result. Continue independent work within the current authorization rather than seeking approval for every shot. Learning or planning does not itself authorize rendering, installing models, scheduling, or publishing. When production is already authorized, continue through the stages the host can actually perform.

Shared decisions

  • Preserve the effect and its causes. When scenery may change, protect the expressive moment, information order, recognizable characters, and relationships needed to make the effect work. Set dressing similarity is not an automatic success criterion. Keep any facts the user explicitly locked.
  • Separate evidence from interpretation. Record body facing, head facing, visible eye direction, and the object being handled separately. A turned head or working hand does not prove where the eyes look. Keep inferred intention and invented direction distinct from observed behavior.
  • Protect the path when the path carries the moment. A glance, wind-up, interrupted action, imbalance, or recovery may matter more than the final pose. Endpoints constrain boundaries; the intermediate performance still needs design and output evidence.
  • Manage attention. Decide who the audience should watch, the cue that transfers attention, and the space to read the reaction. One character may become quieter so another can be noticed. More simultaneous movement is not automatically better.
  • Keep consequential speech with its response. A word, negation, interruption or reply may trigger the defining action. Preserve its source line and listener reaction when relevant. Separate ASR text, captions, actually heard delivery and visible speaker evidence; an anonymous voice label is not a character identity. Silent material needs no invented dialogue.
  • Fit the performer and style. Humans, realistic animals, stylized creatures, and other characters support different expression channels. Do not impose human facial muscle codes, repeated blinks, exaggerated mouths, or continuous motion on every subject. Deliberate stillness can be a performance choice.
  • Keep timing purposeful. Derive timing from the event, not fixed pause lengths, beat counts, joke density, or target runtime. Distinguish requested generation length from the selected edit. Remove redundant waiting while preserving anticipation, reaction, and aftermath.
  • Review what was actually observed. A complete card, prompt, decode, or contact sheet does not prove real-time rhythm or audible quality. Use pending findings where capabilities or evidence stop. Bind approval to the actual version and range reviewed.

This skill supplies planning and review instructions. It includes no media analyzer, generator, renderer, or platform integration. Use tools or other skills already available in the host when the task requires them; no particular backend is required. Tool availability and authorization must both be established before execution. The method cannot guarantee model compliance, audience laughter, or reach.

Source acknowledgments and adaptation limits: sources. They document the method's background; the workflow is self-contained.

Individual skills in this repo

This repo contains 2 individual skills — each has its own dedicated page.

관련 스킬