platform-product-video-harness

[HARNESS INTERNAL — do not invoke directly] Phase 1 of the platform-product-video-harness `generate` workflow. Extracts frames from a reference video (`sample.mp4`) via ffmpeg, reads each PNG via Claude vision, produces a structured JSON outline (scene count, pacing, palette, suggested template) cached at `projects/<slug>/_cache/analysis.json`. Silent — no human gate; output feeds Phase 2 (scriptwriter). Never invoke outside the harness — use /platform-product-video-harness:generate.

[HARNESS INTERNAL — do not invoke directly] Phase 3 of the platform-product-video-harness `generate` workflow. Generates `compositions/<slug>.tsx` from the approved script.md, wires asset references via Remotion `staticFile()`, then runs `npx remotion render` three times (9:16, 1:1, 16:9) to produce silent preview MP4s in `_preview/`. Hosts GATE #2 for human visual approval before Phase 4. Never invoke outside the harness.

[HARNESS INTERNAL — do not invoke directly] Phase 4 of the platform-product-video-harness `generate` workflow. Loads ELEVENLABS_API_KEY from <ProductVideos>/.env (refuses to proceed and prints the remediation block if missing — never auto-writes .env). Calls ElevenLabs TTS per scene + Sound Effects API for an ambient BGM bed. Muxes audio + video via ffmpeg into 3 final MP4s in projects/<slug>/out/. Never invoke outside the harness.

[HARNESS INTERNAL — do not invoke directly] Phase 5 of the platform-product-video-harness `generate` workflow. Read-only verifier. Runs ffprobe on each of the 3 final MP4s to verify aspect ratio, duration (±2s of script target), audio levels (-16 LUFS voiceover, -28 LUFS BGM), no clipped frames, no missing captions, no hardcoded brand colors in the composition. Hosts GATE #3 for final human approval. Never writes any file. Never invoke outside the harness.

[HARNESS INTERNAL — do not invoke directly] Phase 2 of the platform-product-video-harness `generate` workflow. Picks the closest template from skills/script-templates/ based on the analyzer's outline + the user's `brief.md`, fills template variables, drafts `script.md` with scene shot list + voiceover lines + on-screen captions + per-scene durations. Hosts GATE #1 for human approval before Phase 3. Never invoke outside the harness.

9jaLingo TTS API — the secondary voice provider, used by /platform-product-video-harness:translate for Yoruba (yo), Hausa (ha), Igbo (ig), and Nigerian Pidgin (pcm). Covers all four launch-set languages with a single OpenAI-compatible endpoint. Credentials live in <ProductVideos>/.env as NAIJALINGO_API_KEY. ElevenLabs handles English + 28 other languages; 9jaLingo handles the four Nigerian languages; anything else falls back to captions-only.

How to render the same Remotion composition into 9:16, 1:1, and 16:9 in one render pass. The compositor registers 3 <Composition>s per video — same React component, different width/height — and a shared <SafeArea> + scene components adapt to the active aspect ratio at render time via useVideoConfig().

Canonical Q&A protocol for collecting brand kit inputs (name, colors, logo, tagline, fonts) and writing brand/brand.json + brand/logo.svg. Used by /init-video-workspace (Step 4.5, first-time setup) AND by /update-brand-kit (any-time update). Single source of truth so both paths use the same prompts, the same defaults, and the same random palette algorithm.

Pointer to the per-project brand kit canonical source at <ProductVideos>/brand/brand.json. Used by the scriptwriter (personas, banned phrases, tone of voice), compositor (colors, fonts, logo placement), and reviewer (verifies brand tokens are used, not hardcoded). The harness encodes process rules + a neutral starter brand.json template here; the brand itself lives in the ProductVideos repo and evolves with the brand.

How the scriptwriter derives burned-in captions from voiceover lines, syncs them to scene frame ranges, and respects per-platform caption constraints. Captions are mandatory for every voiceover line — Reels / Feed / LinkedIn / Shorts all default to muted playback.

6 hooks across 2 events