Codebase Explainer

Produces an onboarding playlist for a repository: narration-friendly Markdown docs plus NotebookLM explainer videos (one overview + one per subsystem). Claude does the judgement stages (map, author docs); Python scripts do the mechanical stages (preflight, build). Built on notebooklm-py via its CLI.

Outputs (in the target repo)

• docs/onboarding/00-overview.md, NN-<subsystem>.md — analysis docs
• docs/onboarding/videos/*.mp4 — generated videos
• docs/onboarding/index.md — playlist linking each doc → its video
• .claude/idev/codebase-explainer/plan.json — build plan (you author this)
• .claude/idev/codebase-explainer/state.json — resumable progress

Stage 1 — Preflight

Run and read the JSON:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/codebase-explainer/preflight.py"

If ready is false, surface each messages line to the user (install / notebooklm login) and STOP until resolved. Auth is an interactive Google login the user performs once.

Stage 2 — Map the codebase

Read existing idev caches first when present (they are a head start): .claude/idev/smart-context/index.json, .claude/idev/project-map/project.map.md, .claude/idev/architecture-scanner/cache.json. Then dispatch the onboarding-guide agent to identify: tech stack, entry points, and the major subsystems (name, purpose, boundary paths, key files, dependencies, primary data flow). If few/no subsystems are detected (tiny repo), plan an overview video only.

Stage 3 — Author docs + build plan

Write to docs/onboarding/:

• 00-overview.md — architecture, how the pieces fit, main end-to-end flows.
• NN-<subsystem>.md — one per subsystem: purpose, how it works, key files, dependencies, data flow. Write prose that explains concepts, not code dumps.

Then write .claude/idev/codebase-explainer/plan.json:

{
  "style": "whiteboard",
  "videos": [
    {"key": "overview", "title": "Overview",
     "notebook_name": "<repo> — Overview",
     "sources": ["docs/onboarding/00-overview.md", "<1-2 key entry files>"],
     "instructions": "Explain how this system works for a new developer: architecture and main flows.",
     "output": "docs/onboarding/videos/00-overview.mp4",
     "doc": "docs/onboarding/00-overview.md"},
    {"key": "<subsystem-key>", "title": "<Subsystem>",
     "notebook_name": "<repo> — <Subsystem>",
     "sources": ["docs/onboarding/01-<subsystem>.md", "docs/onboarding/00-overview.md", "<key raw file>"],
     "instructions": "Explain the <subsystem> for a new developer: responsibilities, key files, main data flow.",
     "output": "docs/onboarding/videos/01-<subsystem>.mp4",
     "doc": "docs/onboarding/01-<subsystem>.md"}
  ]
}

• sources are file paths (docs + a few key raw files for grounding — hybrid).
• style applies to every video for a consistent look; set it to the --style <visual-style> value from $ARGUMENTS (default whiteboard if omitted).
• If a subsystem positional argument was given in $ARGUMENTS, limit videos to the overview entry plus that one subsystem's entry only (still write the overview doc for context, but omit all other subsystem docs from plan.json).
• If few/no subsystems were detected (tiny repo, same as Stage 2), omit the per-subsystem NN-<subsystem>.md docs and include only the overview video in plan.json.

Stage 4 — CHECKPOINT (mandatory)

STOP. Tell the user the docs are in docs/onboarding/ and ask them to review/edit before any videos are generated. Nothing has been sent to NotebookLM yet. Only continue on explicit approval.

Stage 5 — Build & generate

python3 "${CLAUDE_PLUGIN_ROOT}/skills/codebase-explainer/build_videos.py" \
  --plan .claude/idev/codebase-explainer/plan.json \
  --state .claude/idev/codebase-explainer/state.json \
  --index docs/onboarding/index.md

Generates sequentially (overview first), polling and downloading each video. On a failure (e.g. NotebookLM daily limit) it records progress and stops — tell the user which videos remain and that re-running the same command resumes, skipping completed videos. Report the final playlist at docs/onboarding/index.md.