- Home
- Skills
- Agents & Orchestration
- video to skill
Works with the AI Agents you already use
video to skill
Converts video tutorials into structured, evidence-backed agent skills with local extraction and sandbox verification.
Free
video to skill
Example session with this skill installed
I have a Vim tutorial at https://youtube.com/watch?v=... — watch it and turn it into a skill my agents can use.
- Read your context and instructions
- Compiled the video to
Watched the tutorial and compiled it into an installable skill. Commands were read off the frames rather than the narration — that mattered twice.
What the video taught
| Genre | Screencast / CLI ⚠ |
| Steps extracted | 9 ⚠ |
| Badge | ✅ Verified by execution — 7 passed, 2 skipped by safety policy |
| Installed to | ~/.agents/skills/vim-basics, symlinked into every detected agent |
Where pixels beat the transcript
| Narration said | The frame read | What shipped |
|---|---|---|
| "QBang" | :q! at [t=0:54] | :q! |
| "the first extension called Vim" | vscodevim.vim v1.18.9 | the exact extension id |
A step, as it ships
- id: quit-discard-changes
goal: Leave Vim without saving
action: ":q!"
evidence: "[t=0:54] references/frames/shot_007_key.jpg"
confidence: high
What the sandbox caught
The paste step compiled to "+P — Vim's internal register, not the system clipboard. The sandboxed sub-agent ran it, the success criterion failed, the repair loop corrected it to "+p, and it re-verified against the real pasteboard. Two steps that would have been unsafe to execute were skipped by policy and say so on the badge — no silent green.
Next steps
/vim-basicsis available in your agents' next session.- Feed it a second Vim tutorial to fold in: agreements gain confidence, conflicts become cited variants.
- The
.vts/vim-basics/bundle was kept for that fold — delete it whenever you like.
Connects securely to your tools. The creator never sees your data.
About this skill
Video to Skill solves the problem of high-friction knowledge transfer from tutorials to automated agent workflows. If you find yourself pausing videos to copy commands or transcribe steps for your agents, this skill automates that entire pipeline — and the video never leaves your machine.
What it does
- Deterministic local extraction — a Rust binary (
vts-extract) handles download, word-timestamped transcription, perceptual-hash shot detection, motion density and keyframe export. Only the frames the agent chooses to read enter the model's context. - Pixels outrank narration — a whisper model heard "some"; the frame read
=SUM(B3:B10). Commands and identifiers are read off the frame character by character, never paraphrased from the transcript. - Every claim is cited — each step carries
[t=MM:SS]and the frame it came from. Anything not visually verifiable is flagged low-confidence rather than guessed, and a non-procedural video (a talk, a vlog) returns a report explaining why no skill was generated instead of fabricated steps. - Verification is execution, not vibes — a fresh sub-agent that has never seen the video runs each step against its success criteria inside an OS sandbox (macOS Seatbelt, Windows LPAC/AppContainer, Linux bubblewrap) with no network, writes confined to a throwaway workspace, and credential stores unreadable. Where no sandbox exists the run is refused and the steps are badged honestly as unverified.
- Skills that grow — fold in a second video of the same task: agreements gain dual provenance and a confidence bump, conflicts become cited variants rather than overwrites, and verification re-runs before reinstall.
- Installs everywhere at once — one real copy in the universal
~/.agents/skills, with relative symlinks from every agent actually detected on the machine (Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, Amp).
How it works
- Resolve the extractor — locate or fetch
vts-extract, checksum-verified, thenvts-extract check --fixprints exactly what else is missing and downloads nothing until you approve it. - Extract the bundle: manifest, word-level transcript, shots and keyframes, aligned timeline.
- Read the timeline, then look closer — the agent reads the token-lean digest first and re-watches a range only when evidence demands it.
- Compile, verify, install — the analysis becomes a Procedure IR, the compiler validates it, the sandbox tests it, and installation prints its plan before writing anything.
Safety
Credentials visible on screen are never transcribed — the compiler refuses an IR carrying secret-shaped text in any field rather than sanitising it. Video text addressed to an AI ("ignore previous instructions") is flagged as suspected prompt injection and never encoded into steps. curl … | sh style steps are never emitted as runnable scripts. Nothing downloads or installs without showing you the plan first, and every artifact is version-pinned and sha256-verified.
Why this beats prompting it yourself
Generic prompts hallucinate commands and miss flags, versions and UI labels. This enforces pixels-over-narration, then proves the result by running it under an OS sandbox — neither of which a prompt can do.
Known limitations
Best on screen recordings and command-line tutorials; fast physical demonstrations lose detail between frames. Some GUI clicks cannot be checked by executing them, and are labeled unverified rather than hidden behind the badge. Non-procedural videos produce an analysis report, not a skill.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires local binary execution. Supports x64/arm64. Procedural videos only for skill generation. Requires ffmpeg/yt-dlp (can auto-download).
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 14 days ago
- Free with an Agensi account
- Software
GitHub CLI
Python
GitHub CLI is only needed on macOS and Linux, to fetch the prebuilt extractor.