Skill: video transcript is undiscoverable — it lives in `seo get`, not documented as a transcript source

Start by reading the seo row in packages/twentythree-skills/skills/SKILL.md and the guidance in guide.md. Update those two files to document the subtitle, webinar transcription, and seo get check order, the jq extraction path, and the machine-generated draft caveat. Done means a reader can find and retrieve the transcript without re-transcribing audio.

Summary

An agent (Claude Code) using the twentythree skill failed to retrieve a video's transcript because nothing in the skill points to where the transcript actually lives. The skill documents two transcript-ish sources, both of which return empty even when a transcript exists; the real one is in seo get, described only as "SEO metadata management."

What happened

Goal: pull the transcript for a webinar-recording video (the platform manage page clearly shows a "Video Transcript" / "Full transcription"). The agent reasoned "transcript = subtitles or webinar transcription" and checked exactly those:

• video get <id> -> subtitles_p: false
• webinar transcription list <webinar-id> -> 0 transcriptions
• video subtitle list <id> -> []

All empty, so it concluded no transcript existed and started re-transcribing the audio via Whisper (which also failed first pass -- the 52 MB audio exceeds Groq's 25 MB limit). The transcript was available the whole time via:

twentythree seo get --object-id <id> --json | jq '.data.data.seo_data.transcript'

Path: data.data.seo_data.transcript[], each item {text[], timestamp_begin, timestamp_end}, ordered by start time, flagged seo_data.context.transcript_draft_p: true (machine-generated draft). This is exactly what the manage page renders as "Video Transcript."

Why the skill made this hard

• SKILL.md resource index describes seo as just "SEO metadata management" -- zero hint it embeds the full timestamped transcript.
• There is no "how to get a transcript" guidance anywhere, so an agent naturally tries subtitle / webinar transcription and stops when both are empty.
• The three stores (published caption tracks, live STT transcriptions, SEO auto-draft transcript) are independent, and the only one populated for auto-transcribed videos is the SEO one.

Suggested fix (docs only)

In packages/twentythree-skills/skills/:

1. SKILL.md -- expand the seo row: note it carries the auto-generated timestamped transcript (seo_data.transcript).
2. guide.md -- add a short "Getting a video's transcript" correctness rule with the check order (subtitle -> webinar transcription -> seo get) and the jq extraction one-liner above, plus the caveat that the SEO transcript is a machine-generated draft (not quote-accurate).

Happy to send a PR with these two edits if useful -- flagging as an issue first.