download-video
Skill1 starsChanged 5 months agoMITView on GitHub
Download a video from a URL (YouTube, Instagram reel/post, TikTok, any yt-dlp site) into the vault and turn it into a note. Saves the video + thumbnail under 07-Attachments/, picks the best transcript source (creator subtitles → local Whisper → auto-captions), applies a quality gate, then hands the transcript to the ingest skill. Works with just yt-dlp + ffmpeg; uses Whisper (via uvx, GPU-accelerated when available) for spoken-word videos with no usable captions. Use when the user shares a video link or asks to download, save, or preserve a YouTube/Instagram/TikTok video, reel, or recipe.
- Pipes a download into a shell
- Installs packages
6 other projects use a version of this.
Adopted by
Most recently taken up by
- mofongo/claude-skillstheir own version, first seen Nov 16, 2026
- feiskyer/video-skillstheir own version, first seen Nov 9, 2026
- cyberpinkman/awesome-capturetheir own version, first seen Jul 30, 2026
- biyearly-mesothelioma790/skillstheir own version, first seen Mar 28, 2026
- swyxio/skillstheir own version, first seen Mar 28, 2026
- rameerez/claude-code-startup-skillstheir own version, first seen Feb 4, 2026
Other versions6 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 6
- swyxio/skills★ 199same name
- rameerez/claude-code-startup-skills★ 30same name
- feiskyer/video-skills★ 16same name
- cyberpinkman/awesome-capture★ 2same name
- biyearly-mesothelioma790/skills★ 0same name
- mofongo/claude-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- brainstormingSkill
- grill-meSkill
- ingestSkill
- querySkill
- researchSkill
- setupSkill
- vault-lintSkill
- weekly-reviewSkill
The file
Files
download-video — preserve a web video as vault knowledge
Turn a video URL into vault knowledge: the downloaded video kept in attachments (so it
survives the post being deleted), plus a note built from its transcript and caption. The
heavy lifting — download, metadata, transcript selection, quality gate — runs in
scripts/yt-fetch.sh; this skill drives it and then hands the result to ingest, which
writes the actual source note. Follow the vault conventions in CLAUDE.md (frontmatter, naming,
todo rule, folder icons); do not restate them here.
Folder paths in this skill (07-Attachments/, 04-Resources/, 05-Daily-Notes/, …) are the
template defaults — the authoritative paths live in the "Folder roles" table in CLAUDE.md's
Vault structure section. Resolve folders by role there; never trust the literal defaults on a
renamed or non-English vault.
Dependencies & graceful degradation
Probe with command -v — never assume. The script does this itself and prints platform-specific
install hints when something required is missing.
| Capability | Needs | If absent |
| Download + metadata + captions | yt-dlp, ffmpeg, python3 (hard deps) | Script stops with install hints; nothing else works |
| Whisper transcription (spoken video, no captions) | uvx (ships with uv) | Falls back to auto-captions; if none, transcript is poor |
| Fast Whisper | NVIDIA GPU (auto-detected via nvidia-smi) | Runs on CPU (slower); auto-picks a smaller model |
| Platform | Install hard deps |
| Debian/Ubuntu/WSL | sudo apt install yt-dlp ffmpeg python3 |
| macOS | brew install yt-dlp ffmpeg python |
| Arch | sudo pacman -S yt-dlp ffmpeg python |
| Whisper (any) | curl -LsSf https://astral.sh/uv/install.sh | sh (provides uvx) |
Whisper's first run downloads the model (large-v3 ≈ 3 GB on GPU, small ≈ 0.5 GB on CPU) —
mention this to the user before a long transcription so the wait isn't a surprise.
Transcript source priority & quality gate
The script tries, in order, and stamps transcript_quality in the output frontmatter:
- Creator subtitles →
good— human-authored, trust them. - Whisper (local transcription of the audio) →
good— preferred over auto-captions. - Auto-captions (platform ASR) →
fair— usable but error-prone. - Nothing →
poor— no transcript produced.
Act on the gate after the script returns (it prints RESULT method=… quality=… dir=…):
good→ proceed to ingest normally.fair→ ingest, but treat technical terms, names, quantities, and numbers as low-confidence; keep the[mm:ss]anchors so claims can be checked against the video.poor→ do not ingest blindly. Tell the user, and offer to re-run with--whisper(forces audio transcription) or--cookies-from-browser <browser>if the content was login-gated. Never build a note from a non-existent or garbled transcript.
Steps
- Get the URL and intent. Confirm what the user wants preserved — for a recipe, that's
ingredients + steps; for a talk, the key points. Note the platform: YouTube usually has real
captions; Instagram/TikTok reels usually don't — the recipe often lives in the post
caption (the script captures it as
## Post caption / description) and the spoken steps need Whisper. Resolve the attachments folder by role from CLAUDE.md, then run the fetch script, passing it as the parent dir. The script creates
yt-<id>-<slug>/inside it:bash .claude/skills/download-video/scripts/yt-fetch.sh "<URL>" --out "07-Attachments"Useful flags:
--whisper(force transcription, skip captions),--no-whisper(captions only),--cookies-from-browser firefox(private/login-gated),--model <name>,--force(overwrite). For a typical Instagram reel where you expect no captions, go straight to--whisper.- Read
RESULTand apply the quality gate above. Ifpoor, stop and offer the re-run options instead of ingesting. - Read
<dir>/transcript.md. It carries the metadata frontmatter, the post caption, and the anchored transcript. For a recipe, distil it into clean ingredients and numbered steps — the caption and transcript together, not one or the other. - Hand off to
ingest. Invoke theingestskill on<dir>/transcript.mdso the source note lands in the right resource topic (or area — a recipe belongs in the cooking area, not resources; place by content). Embed both the cover image (![[video.jpg]], the post's hero shot — for a recipe this is usually the finished dish) and the saved video (![[video.<ext>]]), and carry thesource_urlinto the frontmatter. For more pictures (e.g. step-by-step stills), grab frames from the video with ffmpeg:ffmpeg -i <dir>/video.<ext> -vf fps=1/5 <dir>/frame-%02d.jpg(one frame every 5s; tune the rate), eyeball them, and embed only the useful ones. Ingest also handles the daily-note log entry and any follow-up todos per the vault's two-place todo rule. - Confirm what was saved: the note path, where the video lives, and the transcript quality so the user knows how much to trust the text.
Notes
- The raw downloaded files (
video.*,video.info.json, thumbnail) stay in the attachments subfolder — they are the evidence. Don't delete them. - yt-dlp tags and many site URLs are case-sensitive; pass the URL verbatim.
- This skill is intentionally generic (it resolves folders by role and hardcodes no personal paths) so it can be contributed back to the upstream template.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

