danielrosehill/claude-video-editor-plugin/agentic-edit

Agentic Edit

Drive an LLM-based agent over a long video to produce a shorter cut. The agent watches the video (or samples it), proposes timestamped segments worth keeping, and emits a cut-list. The skill optionally renders the cut.

Procedure

1. Resolve backend

Two backends, picked by user preference (or by what's installed):

DATA_DIR="${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/video-editing"
PREFS="$DATA_DIR/preferences.json"
VENV=$(jq -r '.python_venv // empty' "$PREFS" 2>/dev/null)
PYBIN="$VENV/bin/python"
test -x "$PYBIN" || { echo "No plugin venv. Run deps-setup first."; exit 1; }
 
# default backend = video-use if installed
if "$PYBIN" -c 'import video_use' 2>/dev/null; then
  BACKEND="video-use"
elif [ -d "$DATA_DIR/tools/VideoAgent" ]; then
  BACKEND="VideoAgent"
else
  echo "Neither video-use nor VideoAgent is installed. Run deps-setup."; exit 1
fi

2. Inputs

3. Drive the backend

3a. video-use — uses its own conversational loop. Send the prompt + video path; capture the structured cut list.

"$PYBIN" - <<PY
from video_use import VideoAgent  # API may differ — check installed version
agent = VideoAgent(video_path="$SRC")
result = agent.ask("$GOAL — return JSON list of {start, end, reason}")
import json; print(json.dumps(result, indent=2))
PY

Version drift warning. video-use is young and its API is not yet stable. Before assuming the snippet above runs, the skill should "$PYBIN" -c 'import video_use; print(video_use.__version__, dir(video_use))' and adapt. If the import path or method signature differs, surface that to the user and ask them to upgrade or pin a version.

3b. VideoAgent — invoked from its cloned repo:

cd "$DATA_DIR/tools/VideoAgent"
"$PYBIN" -m videoagent.cli --video "$SRC" --task "$GOAL" --output /tmp/cutlist.json

Exact entrypoint depends on the upstream layout — read the repo's README at first use and write the resolved command back to preferences.tools.VideoAgent.entrypoint so it's cached.

4. Validate the cut list

Expect JSON: [{ "start": "00:01:23.4", "end": "00:01:55.0", "reason": "..." }, ...]. Reject and re-prompt if:

• timestamps overlap or run backwards
• any cut extends past the source duration (compare against ffprobe -show_entries format=duration)
• total kept duration exceeds the requested target by >20% (warn but allow)

5. Save the cut-list

mkdir -p "$DATA_DIR/cut-lists"
OUT_JSON="$DATA_DIR/cut-lists/$(basename "$SRC")_$(date +%s).json"
cp /tmp/cutlist.json "$OUT_JSON"

6. (Optional) Render the cut

If mode = cut-list-and-render:

# Build a concat list of trimmed segments
TMPDIR=$(mktemp -d)
i=0
for row in $(jq -c '.[]' "$OUT_JSON"); do
  s=$(jq -r .start <<<"$row"); e=$(jq -r .end <<<"$row")
  ffmpeg -y -ss "$s" -to "$e" -i "$SRC" -c copy "$TMPDIR/seg_$i.mp4"
  echo "file '$TMPDIR/seg_$i.mp4'" >> "$TMPDIR/list.txt"
  i=$((i+1))
done
ffmpeg -f concat -safe 0 -i "$TMPDIR/list.txt" -c copy "$OUT_VIDEO"

Stream-copy by default; if the segments fail to concat losslessly (timestamp gaps, codec quirks), fall back to a re-encode through render-from-library.

7. Confirm

Print:

• backend used + version
• cut-list path + segment count + total kept duration
• (if rendered) output video path, duration, size
• token / time cost reported by the backend, if available

Suggest follow-ups: review the cut-list manually, refine the goal and re-run, or run generate-deliverables on the rendered cut.

Notes

• This skill is a driver, not the agent itself. All semantic decisions come from the backend.
• Treat all proposed cuts as a draft. Always show the cut-list to the user before rendering.
• For repeatable / declarative cuts (no LLM), use editly-render instead.