Recoup Content — Make Video

Every motion job, one skill. Read the bundled references first: references/workspace-context.md, references/account-resolver.md, references/research-context.md, references/content-api.md (the video modes + async create→poll), references/song-sourcing.md (real audio, no placeholders), references/analyze-gate.md (verify before claiming done).

Backbone

Resolve the artist (use the account_id, not row id — wrong one 404s), read context first, layer live signals, generate, analyze-gate (you can't see motion — analyze and regenerate on failure), persist + commit. Real audio only.

Modes

• short — finished 9:16 of the artist + song. The look is a template (bedroom/stage/outside/album-record-store; default artist-caption-bedroom; list via /content/templates). Fire POST /content/create, poll /tasks/runs?runId= every ~10s until terminal, read output.{videoSourceUrl,captionText}. The async path is agent-safe (sync /content/video times out).
• lyric — transcribe for word-level timings (POST /content/transcribe, audio_urls array → segments:[{start,end,text}]), generate/use an on-brand background, burn in synced lyrics, mux full-length. Don't reproduce third-party lyrics.
• visualizer — seamless loop, no text/logos (Spotify rejects them), 9:16, 3–8s; use first-last with identical start/end frames; analyze-gate the seam.
• reformat — edit, don't generate: master → genuinely distinct per-platform cuts (never identical re-uploads — platforms suppress them). Lead 9:16 with the hook (use recoup-song-find-hook for the in-point); captions clear of UI safe zones.

Guardrails

• Verify before delivery — analyze-gate every render; never claim an unverified asset is done.
• Real audio only (references/song-sourcing.md).
• Stop at the asset — no posting.

References

• references/workspace-context.md · references/account-resolver.md · references/research-context.md · references/content-api.md · references/song-sourcing.md · references/analyze-gate.md