segment-topicslisted

# segment-topics Claude-driven topical chunking. No keyword heuristics — Claude judges where the subject changes. ## Inputs - `transcript`: path to `transcribe` output JSON (must have segments or words with timestamps) - `out` (optional): output JSON path. Default: `<dir>/topics.json` next to the transcript. ## Output ```json { "source": "/path/to/source.mp4", "topics": [ { "t0": 0.0, "t1": 87.4, "title": "Intro and channel pitch", "summary": "Greets viewers, teases the compilation." }, { "t0": 87.4, "t1": 213.1, "title": "Soccer challenge fail", "summary": "Tries to juggle a ball, eats the field, screams." } ] } ``` Boundaries are contiguous (`topics[i].t1 == topics[i+1].t0`). Together they cover `[0, duration]`. ## How 1. Build a compact transcript view (timestamped lines). 2. Send to Claude with the instruction "partition this transcript into contiguous topical regions; emit ~5–20 segments depending on length; favor longer regions over choppy ones; each region must be self-contained (one subject, one bit, one anecdote)". 3. Validate: monotonic non-overlapping spans, full coverage, snap edges to the nearest transcript segment boundary so spans align with sentence breaks. ## Why not just include this in pick-segments Topic boundaries are reusable. Other consumers (chapter markers, retrieval, future skills) want the same partition. Keeping it separate also lets `pick-segments` cache-hit even when N or