YouTube SRT → catalog

Two stages. Stage 1 is generic (any channel/video). Stage 2 is the model-catalog pipeline
that produced docs/models/ from @theAIsearch.

Prerequisites

• yt-dlp on PATH (pipx install yt-dlp; if pipx complains about uv: pipx install --backend pip yt-dlp).
• python3 (stdlib only) for stage 2. No API keys — YouTube captions are free.

Stage 1 — fetch subtitles

S=packages/video-transcription/.pi/skills/youtube-srt/scripts
$S/yt-srt.sh -s now-6months "https://www.youtube.com/@theAIsearch"      # channel, date window
$S/yt-srt.sh "https://www.youtube.com/watch?v=VIDEO_ID"                   # single video
$S/yt-srt.sh -o ~/somewhere -l "de.*,de" -n 50 "https://youtube.com/playlist?list=..."

Output per video: <YYYYMMDD>_<id>_<title>.<lang>.srt + .txt ([mm:ss] line, rolling
auto-caption duplicates removed, ~4× smaller — feed THIS to LLMs). index.tsv = id, date,
duration, url, title. Manual subs preferred, auto-captions fallback; when both en-orig and en
exist only en-orig is kept.

Idempotent: .archive.txt skips fetched IDs; IDs whose subtitles failed (HTTP 429) are removed
from the archive and the script prints WARN: N video(s) without subtitles → just rerun later.

Stage 2 — model catalog (docs/models/)

1. Fetch: $S/yt-srt.sh -s <window> "https://www.youtube.com/@theAIsearch".
2. Batch: in <out>/extract/, ls ../*.txt | sort | split -l 6 -a 1 - batch- (~6 videos ≈ 60k tokens per batch; for an incremental update list only the new .txt files and pick unused batch letters).
3. Extract: one general-purpose subagent per batch (host cap ≈ 2 concurrent), prompt:
   Working dir: <out>. Read extract/PROMPT.md and follow it exactly. Batch file extract/batch-X; write extract/X.jsonl.
PROMPT.md holds the schema below; copy it from an existing <out>/extract/PROMPT.md or recreate from this section.
4. Render: python3 $S/build-catalog.py <out>/extract docs/models → README.md, one <category>.md
   per function, local-hardware.md (tiers T0–T5), models.jsonl (merged, one model per line).
   Hand-curated picks.md is never touched by the renderer — refresh it via DocScribe when data changes.

Extraction schema (one JSON object per model per video)

model, family, vendor, category, functions[], open_weights, license, local_runnable, params, size_gb, hardware, tooling, access, strengths, weaknesses, benchmarks, video_id, video_date, timestamp

category ∈ llm coding multimodal-omni image-gen image-edit video-gen video-edit music tts-voice asr-speech realtime-voice 3d world-model avatar-lipsync agent robotics science-medical embedding-other.
Null when the video doesn't say — never invent numbers. Skip sponsor segments.

Merge + tier rules (build-catalog.py)

• Same model across videos = same normalized model string; latest non-null fact wins; category = most frequent; every mention kept as a timestamped source link.
• Hardware tier (local models only): explicit GB → tier; else presenter wording (per ;-clause strongest requirement, cheapest clause wins); else checkpoint size as est.
  T0 phone/CPU · T1 ≤8 GB · T2 ≤16 GB · T3 ≤32 GB · T4 ≤128 GB workstation · T5 multi-GPU.

Pitfalls

• HTTP 429 from YouTube on bulk runs: built-in --sleep-subtitles 2; rerun to fill gaps.
• --break-on-reject makes yt-dlp exit non-zero on a date-bounded run — expected, ignored.
• Auto-captions mishear names ("clawed" = Claude, "Quen" = Qwen); extraction prompt tells the agent to correct obvious ones; residual oddities stay as heard.
• Names that differ between videos ("Qwen 3.6" vs "Qwen3.6 35B-A3B") stay separate rows — normalization is deliberately conservative.
• Specs are the presenter's claims, unverified. The catalog says so; keep that caveat.
• Keep raw transcripts outside the repo; the catalog cites YouTube URLs with &t= only.