screenkite-transcription-cut

Remove silence gaps from a ScreenKite recording based on the mic transcript.

Prerequisites

screenkite CLI at /usr/local/bin/screenkite.
ELEVEN_LABS_API_KEY (underscored) in a .env anywhere up from the bundle.
ffmpeg on PATH.
python3 + requests (pip install requests).

Phase 1 — Open project

/usr/local/bin/screenkite agent project open \
  --path '<absolute-path-to.skbundle>' --json

Confirm the project is current:

/usr/local/bin/screenkite agent project current --json

Phase 2 — Transcribe mic audio

Mic audio lives at <bundle>/media/microphone_*.m4a.

python3 scripts/transcribe_mic.py \
  '<bundle>/media/microphone_<name>.m4a' \
  --edit-dir '<bundle-parent>/<slug>-edit' \
  --language en \        # or zho, auto-detect if omitted
  --num-speakers 1
# Output: <edit-dir>/transcripts/<stem>.json  (word-level, cached)

Script is a thin wrapper around ElevenLabs Scribe. Re-running is a no-op if the JSON already exists.

Phase 3 — Compute silence cut ranges

python3 scripts/compute_silence_cuts.py \
  '<edit-dir>/transcripts/<stem>.json' \
  --min-silence 0.8 \       # seconds — gaps shorter than this are kept
  --pad 0.15 \              # seconds to keep at each edge of the gap
  --output '<edit-dir>/cuts.json'

Inspect cuts.json before applying. It lists [{start, end, gap_s}, ...] sorted by start time. Adjust --min-silence if too many or too few cuts.

Typical values:

0.5s — aggressive, catches short breath pauses
0.8s — balanced default
1.5s — conservative, only obvious dead air

Phase 4 — Dry-run review (mandatory first pass)

python3 scripts/compute_silence_cuts.py \
  '<edit-dir>/transcripts/<stem>.json' \
  --min-silence 0.8 --pad 0.15 \
  --dry-run          # prints human-readable table, no file written

Show the table to the user. Confirm count and longest gaps look right. Do not apply cuts until the user confirms.

Phase 5 — Apply cuts

/usr/local/bin/screenkite agent tool call \
  --name editTimeline \
  --input-json "$(python3 scripts/compute_silence_cuts.py \
      '<edit-dir>/transcripts/<stem>.json' \
      --min-silence 0.8 --pad 0.15 --emit-tool-input)" \
  --json

--emit-tool-input prints {"action":"cut","parameters":{"ranges":[...]}}.

Or pass the saved cuts.json directly:

python3 scripts/apply_cuts.py '<edit-dir>/cuts.json'

apply_cuts.py reads cuts.json, calls the ScreenKite CLI, and prints the updated clip count returned by the tool.

Phase 6 — Verify

/usr/local/bin/screenkite agent tool call \
  --name getProjectState --input-json '{"scope":"summary"}' --json

Check duration shrank by roughly the total gap seconds. Scrub in ScreenKite to spot-check.

If a cut is wrong: editTimeline action=undo (repeat per bad cut).

Anti-patterns

Never run without dry-run first. cut is destructive and undo is one-step-at-a-time.
Don't use the B-roll skill for this. That skill uses setSceneLayout, which does not remove content.
Don't set --min-silence below 0.3s. Sub-300ms gaps are rarely true dead air and cuts there feel choppy.
Don't cut without opening the project first. CLI calls without a current project silently fail.