video-to-skill (opentoolsorg/video-to-skill) + Claude Code / CodexPartly tested🌏 World

video-to-skill turns a YouTube tech talk into an agent skill: every point has an [mm:ss] timestamp and a seen / guessed / unknown evidence label, and the agent loads it when the topic comes up

video-to-skill by developer @prince_twets (prince): give it a YouTube URL or local video, it gets a transcript via summarize CLI → yt-dlp captions → local Whisper, then the agent writes SKILL.md plus segment notes, a cheatsheet, glossary and timestamp jump list—every claim tagged [mm:ss] and [seen]/[guessed]/[unknown]—and validate.py checks the package. His first real test turned a video on the Vercel CEO's production AI-engineering advice into a published vercel-ceo-skill. We ran the environment check and unit tests, validated his published output (pass, 245 evidence labels), and tried fetching YouTube captions from a server (rate-limited, HTTP 429).

Visit site video-to-skill (opentoolsorg/video-to-skill) + Claude Code / Codex ↗
Pricing
Free
Region
Global
Where it's from
Overseas
Added
Oct 28, 2026
Tested on
Oct 28, 2026
Version we used
video-to-skill v2 (opentoolsorg/video-to-skill @ 9cdf066, 2026-10-24); yt-dlp 2026.10.26; Python 3.13
Categories
ResearchCoding
Original author
prince @prince_twets ↗
Our test
Partly tested
Cost / time
The skill is MIT and free, as are summarize, yt-dlp, ffmpeg and faster-whisper. Writing the package uses your own Claude Code / Codex quota; long videos mean long transcripts and more tokens.
Source
X @prince_twets ↗GitHub opentoolsorg/video-to-skill ↗GitHub princepal9120/skills (vercel-ceo-skill) ↗

The task

Keep the methods from talks, tutorials and interviews you've watched so your coding agent applies them on related tasks—with every point traceable to the exact moment in the video, so caption mishearings don't become "facts".

Steps

  1. Install: `git clone https://github.com/opentoolsorg/video-to-skill ~/.claude/skills/video-to-skill` (Codex: `~/.codex/skills/`; the README's princepal9120/video-to-skill URL redirects here).
  2. Dependencies: `npm i -g @steipete/summarize` is recommended (the author says it gets around YouTube's 429 bot checks), plus `yt-dlp` and `ffmpeg`; videos without captions or local files need `pip3 install faster-whisper`.
  3. Check: `python3 scripts/extract.py --check` shows which extraction tier is available.
  4. In Claude Code run `/video-to-skill <video-url> <skill-name>`; or run `python3 scripts/extract.py <url>` alone to get just the transcript and metadata.
  5. The agent reads the transcript, sorts it into frameworks / techniques / demos / anti-patterns, and writes SKILL.md (≤4k tokens) plus segments, cheatsheet, patterns, glossary, timestamps and sources, each claim tagged [mm:ss] and [seen]/[guessed]/[unknown].
  6. Run `python3 scripts/validate.py <skill-dir>`, then read the evidence ledger in sources.md yourself: what was said in the video vs. inferred.
  7. Drop it in your skills folder; the agent loads it from the description on related tasks. Re-watch by timestamp whenever a point looks off.

Prompt

/video-to-skill https://www.youtube.com/watch?v=<video-id> <skill-name> Keep the speaker's own terms; every claim needs [mm:ss] and [seen]/[guessed]/[unknown]; mark likely caption mishearings of names as [guessed] and quote the raw caption; skip sponsor segments; after writing, run validate.py and also check every file under references/ or segments/ for absolute local paths.

What the author says

@prince_twets: he forgot 80% of engineering talks a week later, so he built video-to-skill. Input a YouTube URL or local video; output SKILL.md, segment deep dives, cheatsheet, glossary and a full timestamp jump list; every claim cites [mm:ss] and is labeled [seen]/[guessed]/[unknown]—"auto-captions mishear words; the labels keep mistakes from becoming facts"; validate checks frontmatter, timestamps, evidence labels and strips private paths; extraction falls back summarize CLI → yt-dlp → local Whisper; plain Markdown for Cursor, Claude Code, Codex and others. Workflow: watch → /video-to-skill → review the evidence ledger → drop in skills, in minutes. First real test: the Vercel CEO on production AI engineering → vercel-ceo-skill (invariant contracts, direction audits, adversarial breaker agents), each point linked to its timestamp.

Our test

We did not run the full /video-to-skill in Claude Code, generate our own package, or install summarize CLI / faster-whisper. We shallow-cloned the repo (commit 9cdf066, 2026-10-24), ran `extract.py --check` and the bundled test_vts.py, installed yt-dlp 2026.10.26 in a venv and ran extract.py on the YouTube video from the author's example; then cloned the author's princepal9120/skills, ran validate.py on the published vercel-ceo-skill, and built two deliberately broken copies.

What we found when we tried it:The environment check correctly found only ffmpeg; all 4 unit tests passed. Caption download from our server returned HTTP 429; extract.py didn't crash—it suggested installing faster-whisper and skipped. The author's vercel-ceo-skill (SKILL.md + 12 reference files) PASSED validation, with 215 [seen], 19 [guessed] and 11 [unknown] tags across the package. Stripping timestamps from SKILL.md made it FAIL ("zero timestamp citations"); but a /Users/... private path in references/patterns.md still PASSED and only failed once placed in SKILL.md—so the validator doesn't cover references/.

How to get in

The post is on X and the repo README is in English. It works with any yt-dlp-supported URL or a local file, so it isn't limited to YouTube.

Who it suits

Developers who watch a lot of talks, courses and interviews and want their coding agent to remember the methods; teams that need video knowledge as checkable notes.

Pitfalls

  • In our run, yt-dlp fetching YouTube captions from a cloud server got HTTP 429, and extract.py then reported "no subtitle track and no local speech-to-text" and skipped; on servers install summarize CLI or faster-whisper, or run it on your own machine.
  • In our check, validate.py only scans top-level *.md and segments/*.md; the author's own vercel-ceo-skill keeps its detail under references/, which isn't checked—we planted a /Users/... private path there and it still PASSED.
  • The validator only requires at least one timestamp and one evidence label anywhere in the package, not per claim; "every claim labeled" depends on the agent and your own review.
  • The "Vercel CEO talk" is actually an AI LABS breakdown video of an interview (sources.md says so), not the CEO's own words, and some [seen] tags also lean on operator notes—treat skills built from secondhand videos as secondhand.
  • The repo is very new (~1 star) with one published example; the README's segments/ layout differs from the references/ layout the example uses.
  • Auto-captions mishear names (the example tags claw.md → CLAUDE.md as [guessed]); re-watch by timestamp before relying on key points.

More in this category