A second brain from YouTube.

Bring one channel. Here's which one.

One public channel you actually learn from. The blueprint pulls public captions, so it has to be content that is public or that you have the rights to process. Start small on purpose: the example config sets max_videos: 3, and three videos is enough to see the whole shape. The channel gets a short slug in the config; that slug becomes part of the file paths, so pick one you will not want to rename.

The starter vault in the blueprint folder is a tiny finished example: one waypoint index, one waypoint, one card, one raw transcript. Open it first. It is the shape your own vault will take.

What happens the first time?

Six steps, all on your machine. Your agent does the typing; here is what goes by on the screen so none of it is a stranger.

• It checks Git first. The raw captions are other people's words, so your agent checks that the raw folders are ignored by Git, and adds the line to .gitignore on your yes.
• It checks for the caption puller. yt-dlp, the tool that fetches public captions: brew install yt-dlp on a Mac, winget install --source winget --id yt-dlp.yt-dlp on Windows. Python 3.10 or newer underneath, and nothing else. If build.py says yt-dlp is not found, close and reopen the terminal.
• You name the channel. Copy channels.example.yaml to channels.yaml, put the channel's URL in, keep max_videos at 3.
• It builds the raw layer. build.py pulls the captions for three videos and files each one as a raw transcript note. Nothing is rewritten.
• It writes rough first cards. enrich.py reads the raw notes, writes a rough first card per video, mostly the opening lines, and files each under ten starter waypoints made for AI channels.
• Your agent rewrites them for your channel. It rewrites each card from its transcript, three passes, each denser than the last, and marks it draft. Then it reads all the cards, proposes about ten waypoints that fit this channel, and waits for your yes before it writes them. Then you read, in Obsidian, or just ask your agent to show you the L1 index.

cp pipeline/channels.example.yaml pipeline/channels.yaml
python3 pipeline/build.py --config pipeline/channels.yaml --vault vault --max-videos 3
python3 pipeline/enrich.py --vault vault

What you can ask your agent, from now on

Five asks that work once the vault exists. Copy, paste, read what comes back. They all follow the rule: waypoints first, cards second, raw only for a quote.

Words you now own

• Cairn. A waypoint: one durable note about a thing the channel keeps coming back to. Layer one. About ten of them are the route map.
• Card. One dense note per video, written by your agent from the raw transcript. Layer two. A draft until a person reads it.
• Raw transcript. The captions as pulled, never edited. Layer three. Where a quote gets checked.
• Caption. The text YouTube already has for a video. yt-dlp fetches it; nothing is transcribed here.
• Chain of Density. The way your agent writes the cards: three passes over the transcript, each denser than the last, so the card says more in fewer words.
• Embedding. A number-form of a card that a database can search by meaning. Optional; it is what the Supabase step adds.

Before you build more

Four habits the blueprint asks for, because a second brain is only worth having if you can trust it.

• Rights first. Public content, or content you are allowed to process. Nothing else goes in.
• Raw stays out of Git. Full transcripts of other people's videos stay on your machine. Your agent checks the raw folders are ignored by Git before the first build.
• Read the cards. Generated cards are drafts. A card nobody has read is a guess with a nice heading.
• Secrets stay in Infisical. The optional steps need keys. The example env file names them; the real values never go into the vault, the config, or git.

Fair questions