Turn talks and videos into skills

Durable learning artifacts from ephemeral media:

• Extracted steps: prerequisites, commands, failure modes—separate from hype.
• Source links and timestamps so humans can verify controversial claims.
• Skill-shaped output: when to use it, when not to use it, and how to test.

This complements research synthesis but starts from long-form media instead of papers or tickets.

• Skills give the output a home in-repo rather than a forgotten transcript.
• Iterative passes (outline → extract → verify) match how good notes are made.
• Multi-model routing: cheaper passes for chunking, stronger for synthesis.

1. Capture licensing: may you transcribe and redistribute excerpts?
2. Chunk long videos; summarize each chunk before whole-video synthesis.
3. Separate facts from speaker opinion explicitly.
4. Cross-check any actionable step against docs or running code.
5. Publish as skill + changelog when the source video updates.

• Transcript or captions when possible—better than raw audio for accuracy.
• Topic scope so the skill does not become “everything from one keynote.”

1. Draft outline of the talk from chapter markers or slide titles.
2. Extract procedures only—move stories to an appendix.
3. Encode guardrails: OS versions, deprecated flags, known bugs.
4. Peer review with someone who did the task manually.
5. Dogfood the skill on a real task before wide rollout.

• “List claims that need a primary source beyond this video.”
• “Turn section B into a checklist a junior could follow.”
• “What changed since last year that makes this advice stale?”

• Spot-check every high-stakes step against an official doc.
• Version pin external tools referenced in the skill.
• Attribution block links original creators clearly.

• Conference demos that skip error handling in real systems.
• Hallucinated CLI flags from fuzzy audio.
• Over-long skills nobody reads—split by scenario.
• Copyright issues with pasting long transcript excerpts.