Talking head

Every decision comes from the words. Transcribe first (transcribe on each relevant clip, or reuse a cached semanticProfile.transcript), then cut, frame, caption, insert and mix in that order.

The edit people pay for: a designed insert on each spoken idea, the key word of each caption line restyled, a sound on each element that appears. Cuts stay sparse; the motion is inside the take.

Workflow

1. Understand. Transcribe the speech, read silenceRegions and the measured face boxes. Fragmented words are not evidence of what was said.
2. Structure. Cut to the strongest complete idea: start on a sentence that stands alone (never "so, yeah"), keep the setup needed to understand it, end right after the payoff line. Remove confirmed pauses, restarts and repetition only: trimSilence for measured silences, cutRanges for word-timed filler and retakes. Keep breaths and meaning. Never shorten an answer into a claim the speaker did not make, and keep the full question when the user asks for it.
3. Framing. punchIn marks emphasis or hides a jump cut within one continuous take, not every sentence; leave amount out and it hits as hard as the direction says. A cut between camera setups follows the shot-size rule in film-editor instead. Rotate it with a cutaway. On a vertical canvas use reframeAuto with follow:"faces" so every shot is framed on its measured face.
4. Captions. addCaptions by clipId, 3 to 4 words a line, with emphasis naming the word each line leans on. On vertical the default y of about +0.2 H keeps them above the bottom 22% (platform UI), and never over the face. Skill: captions.
5. Inserts. One designInserts pass after the cut, one per idea, not one per sentence: a said figure counts up, a point lands as a keyword, steps tick. at is the timeline second of the spoken word. Use it, not addTitle, for graphics on what is said. On the line worth stopping for, wordsBehind puts big words behind the speaker, and placeSpeech sets a line as large type around the frame. Read references/inserts.md for which fits.
6. A sound on each pop, a bed under it. Each insert's start in the result gets a short fetchStockMedia mediaType:"audio", audioPurpose:"sound_effect", under the voice; musicBed for the bed (music-bed). A brief that rules them out wins.
7. Voice. Level with setAudio (compress). Hiss, rumble and hum: audio-cleanup.
8. Check. agentWorkspace action review: mode speech across every join and the ending (no word cut, every thought complete), mode audio for the mix, a frame at each insert.

Gotchas