video-to-skill — instructions for Claude

This repo turns a video (YouTube URL or file) into a reusable Claude skill. The user has nothing to prepare or run: you install, extract, orchestrate and assemble. Never ask them to run a command. Talk to the user in their language. Write the generated skill in the language of their request unless they ask otherwise.

When the user gives a video

Triggers: a video URL or file, "turn this video into a skill", "learn from this video", "read the repo and do it", /video-to-skill <URL>.

1. Python — PY = python or py -3 on Windows (never python3, often a fake Microsoft Store shortcut there); python3 on macOS/Linux.
2. Prerequisites — PY scripts/bootstrap.py --check --json (a few seconds). If "ok": false: say so in one sentence, then run PY scripts/bootstrap.py --json --no-skill (2 to 10 min: give it a 15-minute timeout or run it in the background, do not interrupt it). Only if that fails, follow LLM_SETUP.md.
3. Method — follow SKILL.md with SKILL_DIR = the root of this repo. Always run the scripts with a relative path from the root (PY scripts/extract.py …): those are the commands the repo's permission rules allow. Working folder: v2s_work/<short-id>/ (e.g. the YouTube id).
4. Video longer than 45 min — long-video mode of SKILL.md:
   • one general-purpose sub-agent per part, all launched in parallel, with the instruction: "Read and carry out <absolute path>/part_NN/AGENT_PROMPT.md. Write the notes in <the user's language>. Only write the notes file it names. Reply in 3 lines.";
   • then PY scripts/assemble.py v2s_work/<id> --check: a failing part (incomplete visual log, missing notes) → relaunch its sub-agent with the reason;
   • then plan → assemble.py --plan … --out skills_out/<name> → merges → the skill's SKILL.md.
5. Delivery — automatic, no question — the final skill goes to skills_out/<name>/, then install it with PY scripts/install_skill.py skills_out/<name>: it validates the skill, installs it in ~/.claude/skills/<name>/ and backs up any previous version. Don't ask for confirmation (the user chose automatic install). If validation fails, fix the skill and run it again. Only install elsewhere (--project) or skip installing if the user asked for it. The only question you may ask during the whole run is about cookies, if YouTube requires them.
6. Final message, short: skill name, where it is installed (and that deleting that folder removes it), number of methods, coverage (parts, contact sheets reviewed, densified passages), main uncertainties.

Other requests

• Just download: PY scripts/download.py <URL> [--out downloads] (JSON on stdout; on failure, read hint).
• Install / repair the environment: LLM_SETUP.md.
• Change the tool itself: scripts live in scripts/, sub-agent templates in references/. After a change, run PY scripts/bootstrap.py --no-skill --skip-model (its self-test covers extraction, splitting and validation).