Chunk Captions

chunk-captions

Claude reads a clip-local transcript and returns an ordered list of caption chunks. Each chunk is what one breath/clause would naturally hold (typically 3–6 words) and the chunks together cover every word in the transcript exactly once, in order.

Invoke

.claude/skills/chunk-captions/chunk-captions.sh <clip_transcript.json> <out.json>

• clip_transcript: per-clip transcript with words[] (clip-local timestamps)
• out: path for the chunks JSON

Output

{
  "chunks": [
    {
      "text": "I saw a car today",
      "t0": 0.0,
      "t1": 1.42,
      "words": [{"w": "I", "t0": 0.0, "t1": 0.15}, ...]
    },
    {
      "text": "and it was red",
      "t0": 1.42,
      "t1": 2.55,
      "words": [...]
    }
  ]
}

Rules

• Every word in transcript.words appears in exactly one chunk, in source order.
• Chunk t0 / t1 derive from the first/last word in the chunk.
• Chunks are monotonic and non-overlapping.

How

build_prompt.py indexes every word 0..N-1 and asks Claude to return ordered groupings (lists of word indices). parse_reply.py reconstructs each chunk from its word indices, validates full coverage, and emits chunks.json. On malformed/missing reply it falls back to a deterministic 5-word grouping so the pipeline never stalls.