AI agent skill · by @pedronauck

Kb Yt Channel

Create and maintain Karpathy KB topics from YouTube channels through scaffolding, bulk transcript ingestion, validation, and indexing. Excludes single videos, general summaries, and other sources.

Archived 2026-10-23 · Source updated 2026-09-14. Check the original repository for the latest version.

SKILL.md instructions

# KB YouTube Channel

Use the native `kb ingest channel` for channel resolution, transcripts, dedup/resume, retries, and pacing. The bundled `scripts/ingest-channel.py` adds topic organization under `yt-channels/`, metadata, indexes, validation, and a run report.

1. Resolve the vault, channel/playlist URL, topic slug/title/domain, video selection (`--limit N` or `--all`), and transcript policy (`captions`, `auto`, or `stt`) from the request. Ask only for choices that cannot reasonably be inferred.
2. Check required tools once, reusing current evidence: `kb`, `yt-dlp`, `qmd` when indexing, and `ffmpeg` plus the configured provider when STT is requested. Do not install optional networking dependencies as a default preflight.
3. Run `python3 <kb-yt-channel-dir>/scripts/ingest-channel.py` with `--vault`, `--channel-url`, `--topic-slug`, `--title`, `--domain`, selection, and `--transcribe`. Resolve the helper from this skill directory. Native-language captions are the default; override `--sub-langs` only when needed. `--no-index` defers indexing; `--embed` requests embeddings.
4. A preview is optional. The existing `--dry-run` **creates/updates the topic skeleton and metadata**, then lists the ingest plan without transcripts; it is not a read-only preview. Use it only within authorized topic creation/update. Do not require preview followed by a repeated full preflight.
5. Inspect the helper's JSON summary and failures. Reuse its completed topic-info/lint/index validation; run additional checks only for missing evidence. When searchability is part of the task, verify one representative `kb search` in the topic/collection after indexing.
6. Report actual ingested/skipped/failed counts and the report path. Compile wiki articles only if the task includes synthesis; transcript ingestion alone does not imply compilation.

## Failures and references

Start with the existing network configuration and native bounded retries/backoff. On rate limits, honor server guidance, lower concurrency, or increase pacing. Configure authenticated cookies or a permitted proxy only when an observed access/network problem requires it; never expose credentials or mandate a residential proxy for large channels.

Read `references/channel-topic-contract.md` for metadata/output details and `references/troubleshooting.md` for an actual failure. Resume the same request after repair; native dedup preserves successful ingests. Missing captions can justify `auto`/STT if that scope and provider cost are authorized. Preserve partial outputs; use the supported refresh path for a specific transcript rather than deleting user data by default.