/notes

Run the LecturAL pipeline explicitly for: $ARGUMENTS

This command performs heavy work (media download, optional long CPU transcription, OCR). Follow the procedure exactly and never summarize around a failure.

Always invoke the CLI as uv run --project "${CLAUDE_PLUGIN_ROOT}" --extra run lectural .... This uses the plugin's project and its uv.lock, so the Python run dependencies (including the pinned OpenCV/OCR stack) resolve consistently regardless of the current shell/venv. Do NOT use a bare lectural, python -m lectural.cli, or uvx --from "...[run]" — the session Python may lack the extras, and uvx ignores the lockfile and can pull an incompatible OpenCV. ffmpeg must still be on PATH.

1. If $ARGUMENTS contains no YouTube URL, ask the user for the lecture URL and stop until they provide one. Do not run anything yet.
2. Preflight. Run uv run --project "${CLAUDE_PLUGIN_ROOT}" --extra run lectural doctor --fix and interpret the exit code:
   • 0: ready, continue.
   • 2: surface the first missing/incompatible item and its one-line hint, tell the user to run /lectural:setup, then stop.
   • 1: report the doctor output, then stop.
3. Run the pipeline: uv run --project "${CLAUDE_PLUGIN_ROOT}" --extra run lectural $ARGUMENTS. Treat any non-zero exit code as a hard failure; report the coverage failure and stop. Do not work around it with arbitrary summarization.
4. On success, enrich every generated notes.md by following ${CLAUDE_PLUGIN_ROOT}/skills/lectural/references/summary_prompt.md exactly, reading only local run artifacts (synthesis_input.json, transcript.md, and frames/*.png when OCR text is garbled). Use zero external LLM tokens. Remove every unenriched marker and preserve the line-1 notes marker, the seven section anchors and their order, every <a id="sec-N"> anchor, the slide <img> tags, the coverage footer, and the youtu.be citation deeplinks.
5. Report each output folder and confirm the completeness gate passed.