Module 03 · Lesson 04 of 5

The Capture Workflow

8 min read

Most knowledge doesn't start as text. It starts as a video, a talk, a voice memo, a client call. A simple capture workflow closes the gap between "I heard something useful" and "Claude can find it."

A Brain is only as good as what you feed it. The first two layers give Claude a brain and a way to query it. This layer keeps the brain fed.

The gap where knowledge dies

The hardest knowledge to capture isn't in documents · it's in people's heads and in the things you watch and hear. You pick Postgres over Mongo, maybe the decision gets written down. But the reasoning, the tradeoffs, the context that made it obvious · that's lost.

Same with every YouTube talk, conference session, voice memo, and client meeting. The gap between "I watched something useful" and "Claude can find it" is where most knowledge dies. Capture has to be one step, or it won't happen.

The two-step capture workflow

No magic CLI required · two steps, both with tools that exist today. Get a transcript, then let Claude turn it into a structured note in your brain.

Step 1 · Get the transcript.

A call you recorded is already transcribed if you use Fireflies, Otter, or Granola · export the transcript as text.
A YouTube talk · pull the captions locally, no re-watch needed:

yt-dlp --write-auto-subs --skip-download --sub-format vtt "https://youtube.com/watch?v=..."

A local recording or voice memo · transcribe it on your own machine with Whisper:

pip install -U openai-whisper
whisper "/path/to/meeting.mp4" --model small --output_format txt

Step 2 · Let Claude file it into your brain.

Open Claude Code in your vault and paste this, pointing it at the transcript:

Read this transcript and add it to my brain (file path or pasted below).

1. Extract only the durable knowledge: distinct claims, named frameworks or
   mental models, actionable techniques, and concrete examples with context.
   Skip the filler and small talk.
2. Write ONE note into inbox/ with a prose-as-title filename (a falsifiable
   claim, not a topic), frontmatter (source, date, type), and wiki-links to
   related notes already in the vault.
3. Print a 3-line summary of what you captured and where you filed it.

Everything stays on your machine: Whisper transcribes locally, and Claude reads the transcript and writes the note. A transcript in · a structured, linked, retrievable note out.

Capture before the detail evaporates

The highest-value knowledge is verbal · what got said on the call, the reasoning behind a decision, the framework someone dropped mid-talk. Record every call, then run the capture workflow within 24 hours, while you still remember the context. A 90-minute talk you'd never rewatch becomes 12–18 retrievable claims. Make it a reflex: call ends, transcribe, file it.

What you get from a single 90-minute talk:

12 to 18 distinct claims worth preserving
3 to 5 named frameworks or mental models
5 to 8 actionable techniques
2 to 4 concrete examples with context

That's the material that compounds when Claude retrieves it six weeks later on a completely different problem.

Why this matters for meetings

Meetings used to be the ultimate knowledge sink. Now you record every conversation, run the capture workflow on the transcript, and the tacit knowledge locked in people's heads becomes a structured graph node in your brain.

This isn't about meeting summaries nobody reads. It's active synchronisation between your thinking and its externalised representation · the version of itself Claude can actually search.

The self-improving graph

This is the part that changes everything. Agents don't get bored with maintenance. They don't skip the update because they're late for a meeting. The thing that killed every wiki is exactly what agents are built for.

Over time the system:

Notices when two notes contradict each other and flags the tension
Notices when the spec is out of sync with the codebase
Accumulates friction signals automatically
Proposes structural changes when the architecture creates drag

It refactors its own instructions. It evolves its own structure. The Brain stops being something you maintain and starts being something that maintains itself.

The weekly rhythm

Capture is step one. Processing is step two. Set a recurring 20-minute review:

Process the inbox/ · promote captures into curated knowledge, delete the noise
Review your memory files, prune what's stale
Confirm the most valuable recent sources actually got ingested

"The illiterate of the 21st century will not be those who cannot read and write, but those who cannot learn, unlearn, and relearn."
paraphrasing Alvin Toffler

A Brain that ingests, processes, and prunes is a business that learns faster than it forgets.

Start here

Don't try to ingest everything. Run the capture workflow on your last three most valuable recordings · the talks or calls you'd actually want Claude to remember. Then make capture a reflex: every call, every talk, transcribe and file it. The graph compounds from there.

Open your brain

Read this transcript and add it to my brain (file path or pasted below). 1. Extract only the durable knowledge: distinct claims, named frameworks or mental models, actionable techniques, and concrete examples with context. Skip the filler and small talk. 2. Write ONE note into inbox/ with a prose-as-title filename (a falsifiable claim, not a topic), frontmatter (source, date, type), and wiki-links to related notes already in the vault. 3. Print a 3-line summary of what you captured and where you filed it.