cliptalk-cover-director

Produces evidence-backed cover candidates and three reviewable local cover variants for a ClipTalk video. Use when the user asks for a cover, poster frame, or thumbnail; do not use for timeline editin

SKILL.md

GitHub

Cover Director

Create a recognizable, faithful, small-screen-legible cover and require an exact user selection before changing the current cover.

  • Prefer an accepted cut as the source. Fall back to the source video only when no accepted output is available.
  • Propose diverse frames across evidence-backed moments. Preserve source time, evidence references, component scores, and rejection reasons; suppress black and perceptually duplicate frames.
  • Rank with the declared 100-point policy: request alignment 25, subject readability 20, emotion or action 20, visual clarity 15, title-safe space 10, and distinctiveness 10.
  • Render three local, source-derived directions: clean, editorial, and cinematic. Add title text only when the request explicitly supplies it; do not invent a quotation or claim.
  • Keep every output as a review artifact. review_cover_variants must stop for the user to choose one exact variant, even when the surrounding Agent uses autonomous review.
  • confirm_cover may version and activate only that reviewed variant and matching content hash. Preserve previous versions, never modify the video output, and never publish externally.
  • Do not claim generated imagery, identity retouching, platform certification, or formal campaign export. Those capabilities are outside this version.