youtube-chapters-tags

YouTube Chapters and Tags

Stage 4. Two deliverables from one video, run independently so either can be asked for alone.

1. Chapters: MM:SS – Title for the description and YouTube's chapter UI.
2. Tags: a tag string matching the pattern measured across Ben's recent videos.

Precondition: the video has to be Unlisted

A Studio draft has no caption track. No captions means no timestamped transcript means no chapters. YouTube generates captions only after the video is saved at a real visibility and its checks finish. Private does not work either.

So youtube-studio-setup runs first. If the video is still a draft, or Unlisted but not yet captioned, say so plainly and stop. Do not fake chapters from the script, the Notion card, or the outline. Every timestamp has to come from the transcript.

Check for captions before anything else:

const h = await (await fetch('/watch?v=VIDEO_ID')).text();
h.includes('captionTracks')

False means come back later. It usually clears within the same session for a video under 20 minutes.

Workflow A: chapters

1. Pull the timestamped transcript

YouTube is IP-blocked from the sandbox. youtube-transcript-api and curl both fail; do not retry them. Everything goes through Claude in Chrome.

Use the transcript-panel recipe in references/chrome-transcript-extraction.md.

Do not try to fetch the caption track's baseUrl out of captionTracks. It gets blocked, because that URL carries session tokens. Found 2026-08-25. The DOM panel is the working path.

Store the result on window.__ts. JS responses cap at about 1000 characters, so read it back in slices of roughly 8 lines. Tabs also die mid-session and take window with them, so re-derive rather than assuming state survived.

2. Find the transitions cheaply

Reading 100 segments at 8 lines a call is a dozen round trips. Filter first, read only the windows around each candidate:

const p = /(first|second|thirdly|lastly|and last|best practice|now let me|moving on|
             demo|review|recommend|here above|show you)/i;
window.__ts.map((l,i)=>({i,l})).filter(o=>p.test(o.l)).map(o=>o.i+' '+o.l.slice(0,40)).join('\n')