bookend-trim

Picked segments tend to begin or end mid-sentence because pick-segments and verify-coherence choose by topic, not prosody. This skill nudges each span's [t0, t1] to the nearest clean sentence boundary so the resulting short feels like a complete thought.

Invoke

.claude/skills/bookend-trim/bookend-trim.sh <segments.json> <transcript.json> <out.json> [extend=6.0] [dmin=20]

• segments: output of verify-coherence (or pick-segments)
• transcript: full-source transcript (whisper output with segments[])
• out: same shape as input, with t0 / t1 adjusted and a bookend_note
• extend: max seconds Claude may push outward in either direction (default 6.0)
• dmin: min surviving duration; an adjustment that collapses below this is rejected

How

This project's transcribe skill runs whisper.cpp with --max-len 1 --split-on-word, which strips punctuation. So a pure ends-with-./?/! snap finds nothing and inter-word gaps are also collapsed. Instead we delegate to Claude:

1. For each span, build a context window: every whisper transcript line (segments[], each ~5s of speech) within ±extend of [t0, t1].
2. Ask Claude (single batched claude -p call across all spans) to pick a new t0 that starts at a sentence start and a new t1 that lands on a sentence end. New values must be transcript-line boundaries within the window.
3. Cache on input mtimes; out-of-window or too-short adjustments are rejected and the original [t0, t1] is kept.

Per-span bookend_note records Δt0 / Δt1 and Claude's reason. Pairs well with tighten-pace downstream, which removes dead air inside the re-bookended clip.