Company-Wide Second Brain (5 steps)

We built a company-wide second brain for our content and outbound.

Here are the 5 steps to build your own.

First, why this even matters.

Everyone is racing to put AI agents into their company right now. Most people miss the actual reason it's such a big deal.

When an agent does the work, output gets close to free. You can run 10 of them in parallel. So the bottleneck stops being how much you can produce and becomes how good each task is.

That's the whole game. Make every task just 2% better, run thousands of tasks a month, and that 2% compounds into a completely different company. Tiny marginal gains, enormous volume.

So I got obsessed with one question: what makes EVERY agent task better at once? Not a clever prompt for one job. Better context for all of them.

Honestly, at first I was skeptical. The doubt was simple: is this even worth it? We'd be spending real time building the engine instead of doing the actual work.

But it's the car-factory thing. You can build cars by hand, or you can spend the time building the factory that builds the cars. Slower to start. Far higher ROI after.

Let me save you the months of doubt. This is our Oxygen Knowledge repo, in 5 steps:

1) raw/ -> immutable sources Every input lands here and never gets edited. Meeting notes, competitor pages, sales and customer call transcripts, even our Stripe revenue data. The agents read it, they never overwrite it. This is your ground truth.

2) wiki/ -> the agents compile a wiki This is Andrej Karpathy's LLM Wiki pattern (he posted it in April, it went everywhere). The agents read raw/ and write one short page per topic: product, ICP, positioning, every competitor. Cross-linked, in Obsidian, so it becomes a graph you watch grow every week. A rules file (CLAUDE.md) tells every agent how to maintain it.

3) index.md + log.md -> the memory system index.md is the catalog every agent reads first. log.md is an append-only history of every change. This is the difference between a memory and a folder of random notes.

4) qmd -> search Karpathy's pattern stops scaling past ~100 pages without search. So we added qmd: a local search engine over the wiki, fully on-device, plugged straight into the agents. They stop digging through files and just ask the wiki a question. 209 docs indexed, ranked answer back.

5) skills -> the output layer Each task is its own skill (/linkedin-copywriter, /researcher, /outbound-copywriter ...). Every one reads the brain first, so nothing starts from a blank page.

Then the part that surprised me. I stopped opening files and just started asking it questions in plain English:

-> "What are the biggest pain points right now?" It clusters them from every sales and customer call we've transcribed in. Their words, not my guesses. -> "Who is actually our ICP?" Our Stripe data is plugged in, so it answers from who is really paying us, not the ICP we wrote on a slide. -> "Give me 5 outbound angles for this segment." Written from what customers actually said hurt. -> It feeds the content engine from the same place.

That's when it stopped being a wiki and became a real second brain for GTM. Our voice, our customers' pain, and our revenue, in one place I can just talk to.

Same models everyone else has. Ours just remember everything: every doc, every call, every dollar that came in.

One Sunday to set up. Better every day since, while we keep working.

I wrote up the whole thing. Folder structure, the rules file, the qmd setup, the Obsidian config.

If you want it:

[1] Connect with Tim Scheuer [2] Comment "Second Brain" [3] Repost for priority access

and I'll send it over :)