youtube-research

LLM-driven research over a YouTube channel. The skill scripts handle deterministic I/O (listing, fetching, transcribing); the synthesis is done by you (Claude) reading the script outputs.

Inputs (collect before starting)

Always ask the user for:

1. Channel(s) — one or more handles, URLs, or @names (e.g. @veritasium, or @veritasium, @3blue1brown). Multiple channels are encouraged when the user wants comparison or broader coverage.
2. Topic — research focus (e.g. "quantum computing"). Used as the topic-file slug.
3. Specific video (optional) — single URL/ID to scope to one video instead of the channel(s).
4. Output format — markdown (default) or PDF, asked at runtime.
5. How many videos per channel to consider — defaults to newest 50 per channel; ask if the user wants more/less. Be explicit that the limit is per-channel so the user can budget their time.
6. Specific question to answer (optional) — if the user has a concrete question they want the research to answer (e.g. "Which approach is most cost-effective for small teams?"), capture it. When present, this becomes the spine of the synthesis: every section should serve the answer, and the artifact gets a dedicated Answer section up top. If absent, fall back to the broader topic survey.

Don't guess any of these. If the user names multiple channels but not the topic (or vice versa), ask before running anything — multi-channel runs cost N× the metadata fetch time.

Use the AskUserQuestion tool to collect these inputs rather than free-form prose. One call, batched questions (output format, video limit, optional research question), so the user answers everything in one shot. Channels and topic usually arrive in the user's initial message — only ask for what's actually missing. For the optional research question, phrase it so "no specific question, just survey the topic" is one of the choices, so the user isn't forced to invent one.

Workspace + artifact layout