Concept explainer (on camera, graphics-led)
Read references/house-rules.md first; it wins. Run order: video-director section 3. This file only covers what is specific.
Status: partly proven. The rough cut and footage slots were tested on a short take; a full video mixing the user's footage with receipts, reactions and cutaways has not been made yet. Expect to fix layout issues on the first one and record them in troubleshooting.md.
What's specific
- The presenter is the user's real footage. Never a lip-synced photo or a generated face unless the user asks for one by name (then see
references/presenter.md, legacy section). - Source:
fetch-video.sh→enhance-audio.sh(keeps the picture) →rough-cut.mjs→ review sheet →voice.mjs --provider file --file out/rough-cut.wav. Details:references/own-footage.md. - Their words are fixed. Write
script.jsonfrom what was actually said (fix transcription spellings). Research still applies: add receipts for their claims; if a source contradicts them, tell the user instead of changing the edit. - Shot plan: user on camera (
presenter-fullscenes) for the hook (first 5–10 s), the close, and one or two lines where their face sells it. About a fifth to a third of the runtime. Graphics (beatsscenes) for the rest, over their voice. presenter.mjs footage out/rough-cut.mp4cuts each on-camera slot from the footage at the right moment, so lips always match. Alternate punch-ins (1.0×, 1.3×, 1.12×) so repeated shots differ.- Reactions and cutaways never cover the user's face. On presenter scenes use a name/topic chip and at most a small side reaction; save the big ones for graphic scenes.
- Pacing: dense first minute, then calmer than faceless (about 5–8 changes a minute).
Cost and time
Transcription only (cents). No voice cost. Render ≈ 1 min per minute of video plus the footage cuts.
References
own-footage.md · engine-api.md · presenter.md · voice-music.md · troubleshooting.md