# Media Distill > Fetches a YouTube video transcript and creates a zettelkasten note. Use when the user shares a YouTube link and asks to save / summarize / take a note (also works with short commands `w/s/b/a `). Triggers on phrases like 'save this', 'make a note', 'bookmark', 'save for later', 'just export', 'summary'. - **Type:** Skill - **Install:** `agentstack add skill-codexis-claude-media-skills-media-distill` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [codexis](https://agentstack.voostack.com/s/codexis) - **Installs:** 0 - **Category:** [Content & Media](https://agentstack.voostack.com/c/content-and-media) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [codexis](https://github.com/codexis) - **Source:** https://github.com/codexis/claude-media-skills/tree/main/plugins/media-distill/skills/media-distill ## Install ```sh agentstack add skill-codexis-claude-media-skills-media-distill ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # YouTube → Zettelkasten Turns a YouTube video into an atomic Zettelkasten note and stores it under `Video/`. ## When to use - User sends a YouTube URL and asks for a summary / note / recap. - User says "save this", "make a note", "bookmark this", "save for later", "just export" about a video. - You need to distill knowledge from a lecture, podcast, or tutorial. ## Determining status from the user's command Status is derived from the meaning of the message or from a short command prefix before the link. | Command | Synonyms / intent | status | |----------------------|-------------------------------------------------|--------------| | `w `, `s ` | "save", "make a note", "watched" | `watched` | | `b ` | "bookmark", "save for later", "watch later" | `bookmarked` | | `a ` | "archive", "just export", "source", "mentioned" | `archived` | If no command prefix is given — infer from the phrasing. Default: `watched`. ## Configuration Edit these settings directly in this file before using the skill. | Setting | Default | Description | |--------------------|--------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | `VAULT_ROOT` | `.` | Path to your Obsidian vault root (where `Video/`, `Book/`, `Person/`, `Author/` folders live). Can be absolute or relative to the repo root. Example: `~/notes` or `/home/user/vault`. | | `NOTE_LANG` | `transcript` | Language for note content. `transcript` = match the video's transcript language. Or set a fixed locale: `en`, `ru`, `de`, etc. | | `TRANSCRIPT_LANGS` | `ru,en` | Comma-separated preferred transcript languages passed to `--lang`. Order = priority. | | `PYTHON` | `python3` | Python executable to use. Change to `.venv/bin/python3` if using a local venv. | **Current values:** ``` VAULT_ROOT = . NOTE_LANG = transcript TRANSCRIPT_LANGS = ru,en PYTHON = python3 ``` To reconfigure — edit the three lines inside the code block above. No other files need to change. ## Dependencies ```bash pip install -r .claude/skills/media-distill/scripts/requirements.txt ``` ## Handling multiple URLs If the user's message contains several YouTube URLs — process each one sequentially, creating a separate note per URL using the process below. Report intermediate progress to the user. ## Playlists, shorts, live - **Shorts** (`/shorts/`) — treated as regular videos. - **Playlists** (`?list=...`) — not supported. Ask the user for a specific video URL from the playlist. - **Live recording** — treated as a regular video if subtitles are available. For an active live stream the transcript is usually unavailable. ## Process (step by step) ### 1. Fetch transcript and metadata ```bash .claude/skills/media-distill/scripts/youtube_export.py --lang ``` The script prints JSON to stdout: ```json { "video_id": "abc123", "url": "https://www.youtube.com/watch?v=abc123", "transcript": "full text...", "lang": "ru", "title": "Video title", "channel": "Channel name", "upload_date": "2024-03-15", "duration_sec": 1080, "channel_url": "https://www.youtube.com/channel/UC...", "channel_name": "Channel name", "channel_handle": "@handle", "channel_description": "Channel description...", "safe_title": "Video title", "safe_channel_name": "Channel name" } ``` On error — a message in stderr, exit code 1. **The transcript is not saved** to the vault as a separate file — it is only used to generate the note, then discarded. Channel metadata is cached in `scripts/.cache/cache.sqlite` (30-day TTL) — repeated videos from the same channel reuse cached data. The cache file is gitignored. ### 2. Check for filename collisions Before creating `Video/.md`, check whether the file exists. **If the file exists** — ask the user via `AskUserQuestion`: - **Skip** — note already exists, do nothing. - **Overwrite** — recreate the note (manual edits will be lost). - **Add video_id suffix** — create `Video/ ().md`. Same rule applies to `Book/.md`. For `Person/` and `Author/` a collision simply means "file exists" and should not be overwritten (see §3 and §4). ### 3. Check and create the Person note Use `safe_channel_name` from the JSON as both the filename and the wikilink target. Check whether `Person/.md` exists: - **If it exists** — continue, use `[[]]` as the author reference. The Person note is not overwritten (even if `channel_description` has changed). - **If missing** — create it from `.claude/skills/media-distill/templates/person.md`, substituting fields from the JSON. **Placeholders in `person.md`:** | Placeholder | Source (JSON) | |---------------------------|---------------------------------------------------| | `{{channel_handle}}` | `channel_handle` (channel handle, `@name` format) | | `{{channel_url}}` | `channel_url` | | `{{channel_description}}` | `channel_description` | ### 4. Create the Video note **Pick the template by status:** | status | Template | |--------------|-------------------------------------------------------------------------------------------------------------------------------| | `watched` | `.claude/skills/media-distill/templates/video.md` (full distillation) | | `bookmarked` | `.claude/skills/media-distill/templates/video.md` (full distillation — produce it now, even if the user hasn't watched yet) | | `archived` | `.claude/skills/media-distill/templates/video-archived.md` (minimal stub — export book/author notes, skip video distillation) | **Placeholders (both video.md variants):** | Placeholder | Source | |-------------------------|-------------------------------------------------------------------------------| | `{{url}}` | `url` from JSON | | `{{safe_channel_name}}` | `safe_channel_name` from JSON (sanitized — must match the `Person/` filename) | | `{{upload_date}}` | `upload_date` from JSON (video publish date, `YYYY-MM-DD`) | | `{{save_date}}` | today's date in `YYYY-MM-DD` | | `{{status}}` | `watched` / `bookmarked` / `archived` (from user command) | | `{{duration_sec}}` | `duration_sec` from JSON | | `{{lang}}` | `lang` from JSON (transcript language) | | `{{title}}` | `title` from JSON (original title used as the note heading) | **Filename:** `Video/.md` (field `safe_title` from the JSON). **Requirements for full-distillation notes (`watched` / `bookmarked`):** - Note language = determined by `NOTE_LANG` (see Configuration). If `transcript` — use the video's transcript language (`lang` field). If a fixed locale is set — write the entire note in that language regardless of the transcript. - Each idea is atomic: one section = one thought. - No retelling — only distillation of meaning. - Tags as `#word`, wikilinks as `[[concept]]`. - Forward-wikilinks on books/authors (`[[Book Title]]`, `[[Author Name]]`) in the body are the trigger for step 5. Write them in the source language as they sound in the video — do not try to guess the canonical English title. Alias resolution happens in step 5. **Requirements for `archived` notes:** - 1–2 sentences of context — **why** this note exists as a source (what is referenced, what it links to). - No "Key ideas", "Practice", "Related" sections — the note is a stub for backlinks. - Primary purpose: capture book/author references so step 5 runs and Book/Author notes are saved. - If the video mentions books: populate the **"Books mentioned"** section — one line per book in the format `[[Title]] — [[Author Name]]`. Write the title and author as they appear in the video (source language); step 5 resolves them to canonical filenames. If no books — omit the section entirely. ### 5. Post-process — book/author references **Trigger:** any forward `[[...]]` wikilink in the Video note body (context sentence for `archived`, body / Related section for `watched` / `bookmarked`) that points to a book or a book's author. **Always run this step when such wikilinks exist** — book notes are saved unconditionally. **Action:** for each such wikilink pass through `book-flow.md`, passing the tentative name (exactly as written in the wikilink) as the input. Several books in one video → one pass per book. **Wikilink rewrite.** After `book-flow.md` finishes and the canonical Book/Author filenames are known, rewrite each matching wikilink in the Video note to Obsidian pipe syntax `[[canonical|tentative]]` — e.g. `[[La Cinquième Discipline]]` → `[[The Fifth Discipline|La Cinquième Discipline]]`. Reason: Obsidian does not auto-resolve `[[alias]]` through the `aliases:` frontmatter field inside wikilinks; the pipe form is required for the link to navigate to the canonical file while keeping the source-language display. If `tentative == canonical`, leave the wikilink as is. ## Common errors | Error | Cause | Resolution | |-------------------------------|----------------------------------------|---------------------------------------------------| | `403 Forbidden` | YouTube is blocking the IP | Run locally, not in CI/cloud | | `TranscriptsDisabled` | Author disabled captions | Transcript unavailable | | `NoTranscriptFound` | No captions for the requested language | Add `--lang en` | | Empty transcript | Private video or no captions | Check video availability | | `Sign in to confirm your age` | Age-restricted | Inform the user — the script does not bypass auth | ## Skill files ``` media-distill/ ├── SKILL.md ← this file (video flow, triggers, common errors) ├── book-flow.md ← book recommendations flow (templates, list/null fields) ├── templates/ │ ├── person.md ← template for Person/ │ ├── video.md ← template for Video/ (watched / bookmarked) │ ├── video-archived.md ← template for Video/ with status: archived │ ├── author.md ← template for Author/ │ └── book.md ← template for Book/ └── scripts/ ├── youtube_export.py ← fetches video/channel data ├── book_lookup.py ← looks up book metadata ├── cache.py ← SQLite cache for channel metadata (30-day TTL) ├── utils.py ← sanitize_filename() — shared helper ├── requirements.txt ← dependencies └── .cache/ ← cache.sqlite (runtime, gitignored) ``` ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [codexis](https://github.com/codexis) - **Source:** [codexis/claude-media-skills](https://github.com/codexis/claude-media-skills) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-codexis-claude-media-skills-media-distill - Seller: https://agentstack.voostack.com/s/codexis - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.