clips

You select from the transcript, not from waveform peaks or a virality score: a clip earns its place when its first line stands alone and its payoff lands before it ends. Editing follows tested tools (auto-editor, the ffmpeg recipes, the caption script), never hand-written frame-processing code.

1. Locate the recording and probe it: resolution, layout (one speaker, two side by side, camera-switched, screen share), duration, loudness.
2. Confirm only what is missing: how many clips, target platforms, audience and topic focus, names to spell right, topics to avoid. Defaults: 10 clips, 9:16, 20–40 s.
3. remember(["show name and hosts", "caption style", "clip length preferences", "topics to avoid", "clips that performed"]).

Phase 2 — Transcript

Extract audio and transcribe once with word timestamps (whisper-ctranslate2 ... --model medium --word_timestamps True --output_format json; pass host and guest names with --initial_prompt). Save to ./video-out/<slug>/captions/full.json and a readable transcript.md with timestamps every sentence.

Phase 3 — Selection

1. Read the whole transcript. List 2–3× more candidates than requested.
2. A candidate opens on a self-contained line (a claim, a surprising number, a question, a contrarian take) and ends after its payoff sentence. Never end on a setup; never open on context the viewer lacks ("as I said earlier").
3. Score each 1–5 on hook strength, standalone clarity, payoff, and quotability; drop anything under 12/20.
4. Write clips.md: rank, start–end, length, first line, payoff line, score with one-line reasoning, suggested on-screen hook (≤8 words), suggested title.
5. Show clips.md and wait for the user's picks.

Phase 4 — Render each pick

Follow the video-editing skill and its recipe reference per clip: cut with 0.2–0.4 s padding at both ends; tighten internal silences with auto-editor (--margin 0.15s); reframe by layout (face-centred crop for one speaker, stacked halves for two side-by-side speakers, blurred fill for screen share); two-pass loudness to -14 LUFS. Then the video-captions pipeline on the clip: re-transcribe the cut clip, proofread names, render with the video-captions command in pop (captions at --y 0.5 on stacked layouts) with the --hook text, burn. Render clips one by one and report progress; save as cuts/clip-NN-9x16.mp4.

Phase 5 — Deliver

Encode per video-spec-sheet; render one cover per clip with video-thumbnail render <frame> --format shorts --layout top --text "<hook text>" from the hook frame (pull it with video-thumbnail frames or ffmpeg -ss); write meta/clips-manifest.md (file, source timestamps, hook text, title, suggested caption copy, any disclosure). QA frames at the hook and the payoff of every clip, and each cover, with read_media_file; if an image can't be read, list the frame files for the user and say placement is unverified.

Skills

Memory protocol

Before starting: remember() as in Phase 1. After completing: memorize() — clip types the user picked or rejected, hook text that performed, layout fixes per show.

Do:

• Transcribe once with word timestamps; select from the transcript.
• Write clips.md and get picks before rendering.
• Re-transcribe each cut clip for its captions; proofread names.
• Save everything under ./video-out/<slug>/.
• remember() before, memorize() after.