choumays/claude-skills

video-to-skill

The pipeline: find videos → watch each one twice (transcript+frames, then Gemini) → merge into one knowledge base → write a skill and/or an agent → test it on a real task.

Work in a folder such as learn/<skill-name>/; everything below lands there.

1. Pin down the target (ask only what you can't infer)

• Role or skill: "short-form video editor", "Facebook ads buyer", "bread baker".
• Output: a skill (instructions any session can load), an agent (a specialist persona with its own tools that loads those skills), or both. A role that bundles several skills ("social media manager") → both. One technique → a skill.
• Where it lives: project (.claude/skills/, .claude/agents/), personal (~/.claude/skills/, ~/.claude/agents/), or this repo as a plugin.
• Sources: the user's links. If they gave none or only one for a broad skill, find 3–8 strong ones (see step 2).

2. Collect sources

Prefer: practitioners showing their real work, recent uploads for fast-moving tools, and videos with on-screen detail. Avoid: hype with no demonstration, reuploads.

# search YouTube without a browser (titles, channel, length, views, url)
yt-dlp "ytsearch15:<topic> tutorial" --flat-playlist \
  --print "%(title)s | %(channel)s | %(duration_string)s | %(view_count)s | %(url)s"

WebSearch also works. Keep a sources.md list with why each one was chosen.

3. Watch each video, two ways

For each source (run independent videos in parallel subagents when there are several):

python3 ${CLAUDE_PLUGIN_ROOT}/skills/watch/scripts/watch.py "<url>" --out learn/<skill>/videos/<id> --scene
python3 ${CLAUDE_PLUGIN_ROOT}/skills/gemini-watch/scripts/gemini_watch.py "<url>" --mode skill \
  --out learn/<skill>/videos/<id>/gemini-skill.md

(Outside a plugin, the scripts are in the sibling skill folders ../watch/scripts/ and ../gemini-watch/scripts/.)

• Use the frame mode that fits the video (see the watch skill's table). For editing or other visual crafts, add --mode visual to Gemini too.
• If one route fails (no key, YouTube blocked, no captions) continue with the other and note the gap. If both fail, stop and tell the user what's needed.
• Then read index.md, open the frames that matter, and write videos/<id>/notes.md using the video notes template in references/templates.md. Every concrete value gets a timestamp.

4. Merge into one knowledge base

Write learn/<skill>/knowledge.md:

• Consensus: what several sources agree on → becomes the default procedure.
• Variants: where sources differ, record both with the conditions under which each wins ("A for talking heads, B for product demos").
• Specifics: exact settings, numbers, shortcuts, templates, prompts, checklists.
• Judgment: the principles behind the steps, and what "good" looks like.
• Unverified: single-source claims, promotional claims, things that may be outdated. Check software claims against current docs when you can.

Leave out filler, sponsor reads, and anything you can't tie to a source.

5. Write the skill

Follow the SKILL.md template in references/templates.md. Rules:

• name: lowercase-hyphenated. description: what it does AND when to trigger it, with the words a user would actually say. This line decides whether the skill fires.
• Body under ~300 lines, written as instructions to Claude ("Do X. Check Y."), not as a summary of videos. Procedure first, then decision rules, then pitfalls, then a quality checklist.
• Push long material (full setting tables, prompt libraries, per-platform specs) into references/*.md and say in the body when to read each one.
• Add scripts/ only for steps that are deterministic and repeated (an ffmpeg export preset, a caption formatter). Test every script you add.
• End with a Sources section: video title, channel, URL, date studied.

6. Write the agent (if requested)

Follow the agent template in references/templates.md, saved to .claude/agents/<role>.md (or agents/<role>.md inside a plugin). The agent file holds the role, judgment and workflow; the skills hold the know-how. List the skills it should use in the skills: field and restrict tools: to what the role needs. One agent can draw on several video-built skills (a "social media manager" = scripting + editing + posting-strategy skills).

7. Test, then tighten

1. Write 2–3 realistic tasks the user would actually give this skill/agent.
2. Run each with a fresh subagent that has only the new skill/agent (not your notes).
3. Compare the output against the videos' quality bar. Where it went wrong, the skill was vague or missing a step: fix the skill text, not the test.
4. Report to the user: what was learned, from which videos, where the files are, how to invoke it, what's unverified, and any setup it needs.

Ethics and limits

• Distill technique in your own words with attribution; don't paste transcripts wholesale into a skill or republish a creator's paid material.
• A skill encodes what the videos show. Flag where real-world practice, a license, or safety training is needed (electrical, medical, financial, legal).