course-digest

Pre-computed context

course-extraction deps: !node -e "const fs=require('fs'),path=require('path'),p=process.env.CLAUDE_PLUGIN_DATA;process.stdout.write(p&&fs.existsSync(path.join(p,'node_modules','@melodic','video-digestion'))?'installed':'MISSING - run setup-deps.mjs (see Prerequisites)')" Playwright Chromium: !node -e "const fs=require('fs'),path=require('path');const b=process.env.PLAYWRIGHT_BROWSERS_PATH||(process.env.CLAUDE_PLUGIN_DATA&&path.join(process.env.CLAUDE_PLUGIN_DATA,'ms-playwright'));const ok=b&&fs.existsSync(b)&&fs.readdirSync(b).some(n=>n.startsWith('chromium'));process.stdout.write(ok?'installed':'MISSING - run setup-deps.mjs (see Prerequisites)')" ffmpeg: !ffmpeg -version 2>/dev/null | head -1 || echo "MISSING — install ffmpeg (see Prerequisites)" ImageMagick: !magick -version 2>/dev/null | head -1 || echo "MISSING — install ImageMagick 7 (see Prerequisites)"

Course Digest

Transform online video courses into structured knowledge and actionable repo recommendations — without watching a single video.

How it works

Uses browser automation (claude-in-chrome) to navigate course platforms, extract transcripts, capture screenshots of code/slides, and collect downloadable resources. Synthesizes raw content into analysis focused on what applies to the current repository.

Where this skill says "deeper research," use whatever external-research capability your project provides. Repo conventions override course claims — surface convention conflicts explicitly; never silently adopt a course's shortcut over team rules.

Emit checklist

For any course digest run (multi-phase content acquisition + distillation + repo-applicability analysis), copy templates/checklist.md into .work/<slug>/course-digest-checklist.md. Tick each phase as completed.

This skill's .work/ root is formally carved out of the marketplace topic-docs convention (https://raw.githubusercontent.com/melodic-software/claude-code-plugins/main/docs/conventions/topic-docs/README.md): the work root resolves through the knowledge plugin's own library_dir seam, not the concern file's memory_dir; slug conformance is form-only (kebab-case [a-z0-9-], ≤ 40 chars, Windows-reserved base names take an -x suffix); and nested <epic>/<slug>/ sub-slices are sanctioned. The root still self-ignores (a .gitignore containing *) and is never committed.

Prerequisites (verify before starting)

1. course-extraction deps — node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/setup-deps.mjs". Installs the pipeline's node dependencies into ${CLAUDE_PLUGIN_DATA} (persists across plugin updates) and provisions Playwright's Chromium into ${CLAUDE_PLUGIN_DATA}/ms-playwright. Idempotent — safe to re-run, and re-run after a plugin update.
2. Platform auth — set COURSE_EMAIL/COURSE_PASSWORD (Dometrain → Clerk) or TEACHABLE_EMAIL/TEACHABLE_PASSWORD (Teachable) in your shell before invoking; they inherit into the pipeline's node subprocess. The env-var prefix is course-config-driven via platformConfig.authEnvPrefix. Session cookies persist under ${CLAUDE_PLUGIN_DATA}/auth/<platform>.auth-state.json and are reused across runs. Interactive manual login is the fallback when no credentials are set — it opens a browser window for you to log in. NOTE: the manual-login prompt (node:readline + headed browser) may not function under headless plugin execution; env-var + cookie-reuse carry the skill regardless, and manual login is a known limitation there, not a blocker. These credentials stay in shell env (rather than migrating to plugin userConfig like this plugin's non-secret scalars) deliberately: a sensitive: true userConfig option still persists as plaintext in ~/.claude/.credentials.json on Windows, so they remain in env until keychain-backed sensitive storage lands there.
3. ffmpeg — required for video frame extraction (scene detection, interval capture). Check: ffmpeg -version. Install: winget install Gyan.FFmpeg (Windows), brew install ffmpeg (macOS), sudo apt install ffmpeg (Linux). Floor 7.1+ (newer codecs — AV1, Opus — degrade or fail below this).
4. ImageMagick 7 — required by classify-frames.js for contact sheet generation (magick montage). Check: magick -version. Install: winget install ImageMagick.ImageMagick (Windows), brew install imagemagick (macOS), sudo apt install imagemagick (Linux). Ubuntu <26.04 ships v6 — v7 may require building from source.
5. claude-in-chrome MCP — needed for Phase 1 (course discovery) and frame extraction HLS URL retrieval. Run tabs_context_mcp as preflight.

If any prerequisite fails, stop and inform the user. Re-run setup-deps.mjs for the node dependencies + Chromium; the media binaries (ffmpeg, ImageMagick) are OS-level installs via your platform's package manager per the commands above.

Running the pipeline scripts

Every extraction script runs through the launcher, which resolves the vendored node dependencies from ${CLAUDE_PLUGIN_DATA} and pins Playwright's browser path:

node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/run.mjs" <script.js> [args…]

Gate on setup-deps.mjs first (Prerequisites above).

Consolidated extraction (Phases 1-2)

The Playwright batch script handles transcripts, video frame extraction, and course metadata in a single run:

# Full extraction: transcripts + frames + metadata
node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/run.mjs" extract-course.js --course-dir <path-to-course-data> --extract-frames

# Transcripts only (faster, no ffmpeg needed)
node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/run.mjs" extract-course.js --course-dir <path-to-course-data>

# Frames only (skip transcripts already extracted)
node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/run.mjs" extract-course.js --course-dir <path-to-course-data> --extract-frames --skip-transcripts

# Course metadata only
node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/run.mjs" extract-course.js --course-dir <path-to-course-data> --metadata-only

Script uses Playwright's bundled Chromium with a fresh temp context (not Chrome itself — Chrome 136+ blocks CDP on default profiles). Auth handled via addCookies() after context launch: with credentials set it logs in and saves state; subsequent runs inject cached cookies automatically. Skips already-extracted lessons (crash-safe, resumable). Runs headless by default (--show-browser to show the browser).

Key technical details:

• Uses --disable-blink-features=AutomationControlled to bypass automation detection on course platform logins
• Chrome profiles are NOT used — Playwright's own Chromium with a dedicated temp context dir in the OS temp directory
• Auth state saved to ${CLAUDE_PLUGIN_DATA}/auth/<platform>.auth-state.json (session expiry depends on the platform's auth provider, configured in platformConfig.authWarnDays)
• HLS video URLs captured via DOM read from the video player element (selector from platformConfig.videoPlayerSelector) — no claude-in-chrome needed
• Frame extraction via ffmpeg scene detection (threshold 0.1), with interval fallback for sparse results
• Content type derived from extraction results (scene detection = code, interval + high dup = talking head) — no manual tagging
• Progress tracking with per-lesson elapsed time, ETA, structured run-report.json output
• SIGINT handler saves progress on Ctrl+C — no work lost on interruption

Long-running extractions (50+ lessons): the CC Bash tool has a hard 10-minute timeout limit. For large courses:

# Run in background with nohup — no timeout limit
nohup node "${CLAUDE_PLUGIN_ROOT}/skills/course-digest/extraction/run.mjs" extract-course.js --course-dir <path> > extraction.log 2>&1 &
echo $!  # save PID

# Monitor progress periodically
tail -20 extraction.log

# Check run report after completion
cat <course-dir>/run-report.json

Platform detection and adapter architecture

Extraction pipeline uses a provider adapter pattern. Platform-specific code lives in adapter modules (extraction/adapters/{platform}.js). The orchestrator (extract-course.js) delegates to adapters via a contract — zero platform-specific code in the orchestrator.

Architecture:

extraction/
  adapters/
    adapter-contract.js       # JSDoc interface + validation + factory
    dometrain.js              # Thin adapter: composes mux (player) + clerk (auth)
    teachable.js              # Thin adapter: composes hotmart (player) + teachable-sso (auth)
  lib/
    players/
      hotmart.js              # Hotmart HLS: intercept, iframe subtitle fetch, canvas frames
      mux.js                  # Mux: DOM src read from mux-player element
    auth/
      clerk.js                # Clerk two-step login flow
      teachable-sso.js        # Teachable simple form login flow
    auth-store.js             # Resolves per-platform auth-state path under ${CLAUDE_PLUGIN_DATA}
    browser.js                # Shared Playwright launch, cookie injection, auth age check
    validators.js             # extraction validators
    config.js                 # platformConfig validation, adapter resolution
  # shared kernel + transcript: @melodic/video-digestion (vendored plugin-wide under ../../vendor/)
  extract-course.js           # Orchestrator only — delegates to adapter + lib
  discover-resources.js       # Discovery tool — uses adapter.detectResources()
  download-resources.js       # Download lesson resources (ZIPs, PDFs) from resources.json
  build-course-json.js        # Teachable-specific: scaffold course.json from curriculum page
  classify-frames.js          # Frame classification (provider-agnostic, no adapter needed)
  generate-manifests.js       # Manifest generation (provider-agnostic, no adapter needed)
  utils.js                    # Provider-agnostic utilities only

Adapters are thin composition layers — delegate to shared lib/players/ and lib/auth/ modules for reusable tech-layer concerns, keeping only platform-specific DOM selectors, URL patterns, and orchestration logic.

Every adapter method returns Result (ok/fail from @melodic/video-digestion/shared/result) — explicit success/failure with timing, operation name, context. No silent catches, no null returns.

URL pattern	Platform	Adapter	Reference
dometrain.com	Dometrain	adapters/dometrain.js	reference/adapters/dometrain.md
*.teachable.com, courses.*.tech	Teachable (Hotmart video)	adapters/teachable.js	reference/adapters/teachable.md
Single public YouTube videos	—	use /knowledge:youtube-digest	—
pluralsight.com	Pluralsight	(future)	—
udemy.com	Udemy	(future)	—

To add a new platform adapter: follow Provider Discovery Checklist to explore the platform systematically, then create adapters/{platform}.js implementing the 5 required methods (extractTranscript, extractHlsUrl, detectResources, deriveLandingUrl, buildLessonUrl). Compose from shared modules where the tech stack matches — e.g., a new platform using Hotmart video + Clerk auth would import * as hotmart from "../lib/players/hotmart.js" and import * as clerk from "../lib/auth/clerk.js", then delegate player/auth methods while implementing only platform-specific DOM selectors and URL patterns. Add the platform to course.json platform field; the factory auto-discovers it via dynamic import. The discovery checklist also serves as regression guide when existing adapters break.

The pipeline

Follow the 8-phase workflow in context/workflow.md, each building on the previous — Discover → Extract → Process Frames → Analyze Code Repo → Validate → Synthesize → Analyze → Recommend (phases 1, 2, 2b, 2c, 2d, 3, 4, 5). Phases 1-2d are extraction (browser + CLI); 3-5 are analysis (LLM-heavy, parallelizable across modules). Storage runs continuously throughout.

Critical rule: ALL context (transcripts + frames + code repo) must be gathered before Phase 3. Module summaries note their context level: [transcript-only], [transcript+frames], [full-context].

Phase 3 modalities ([full-context] requires all three)

Synthesis combines three modalities — transcript (transcript.md), visual frames (PNG files + manifest.json), and companion code (code/repo/<section>/). Each module gets parallel agents (transcript + visual + code exploration), then a synthesis pass. Full modality table + multi-agent approach: context/workflow.md Phase 3.

Multi-modal extraction gaps beyond these three (code OCR from frames, slide-text extraction, audio re-transcription) are evaluated with priority and effort in context/multimodal-evaluation.md.

Dedup semantics

Dedup phase (classify-frames.js --phase dedup) reports near-duplicates but does NOT delete frame files. Actual frame curation happens in generate-manifests.js which sets keep: true/false per frame. Frame PNG files remain on disk — manifests define which frames to use during synthesis.

Phase tracking

course.json includes a phases object recording which pipeline phases have completed and when. Each phase is null (not started) or an object with completedAt timestamp and phase-specific metrics. On resume, check which phases are non-null to skip completed work.

Repo freshness caveat

Companion GitHub repos may be updated after course publication; classify transcript/code discrepancies per context/workflow.md "Freshness verification". All course action items require external research verification before adoption — course content is a starting point for research, not a final answer.

Invocation patterns

Skill supports different scopes:

User says	Action
/knowledge:course-digest <url>	Full pipeline — discover + extract all + analyze
/knowledge:course-digest extract <url>	Phases 1-2 only — extract raw content, skip analysis
/knowledge:course-digest analyze <slug>	Phases 3-5 only — analyze previously extracted content
/knowledge:course-digest status	Show all digested courses and their completion state
/knowledge:course-digest resume <slug>	Resume extraction from where it left off (reads course.json phases)
/knowledge:course-digest continue <slug>	Continue from a prior session — reads continuation prompt file
/knowledge:course-digest (no args)	Auto-detect: check for in-progress courses, resume the most recent

Session handoff protocol

When context is getting large (>40% used) or a session ending mid-pipeline:

1. Write a continuation prompt to <course-dir>/continuation-prompt.md
2. Include: what was completed, what remains, task-by-task breakdown, known issues, quality notes
3. Update course.json phase markers with timestamps
4. Tell the user: "Session state saved. Start a new session and run /knowledge:course-digest continue <slug> to pick up where we left off."

The continue action reads the continuation prompt and reconstructs task context. No args defaults to checking course.json for the most recent in-progress course.

Pacing and user interaction

Courses can have 60+ lessons. Processing all in one session may hit context limits.

• After discovering course structure (Phase 1): present module/lesson list and ask user which modules to process, or confirm "all"
• After every 5 lessons extracted: report progress ("Extracted 5/67 lessons. Continuing...")
• After each module completes: save progress immediately — a crash shouldn't lose work
• If context is getting large (>50% used): suggest saving progress and resuming in a new session with /knowledge:course-digest resume <slug>

Analysis output format

Repo-applicability analysis follows the template in reference/analysis-template.md. Key deliverables:

• repo-candidates.md — Specific patterns/practices from the course that could improve the repository, with references to where in the course they're taught
• action-items.md — Concrete next steps: rule candidates, skill suggestions, architecture patterns, testing practices, work-item candidates

Storage

Generated course output lands under the invoking project's library_dir seam (or ${CLAUDE_PLUGIN_DATA} when no library dir is configured), one self-contained directory per course slug. See context/storage-schema.md for the full directory structure.

Critical rules:

• No video or audio files — transcripts and screenshots capture the content
• Screenshots are PNG files — keep small (resize to 1280px wide max)
• Save progress incrementally — never buffer an entire course in memory
• Each course is self-contained under its own slug directory