The 2-Person Enterprise

Capture · Brain · Act

An agent is a language model that gets context and can act in your systems. Under the hood it's a few markdown files: instructions, context, and a link to your knowledge. No magic.

Every agent we run stands on the same three layers, in this order.

The middle layer is the one that makes the difference. Without it you have a pile of documents. Part 2 is where you build it.

Build your second brain

This is where we started. The pattern is the LLM Wiki, by Andrej Karpathy: Claude writes and maintains the wiki, you steer. One folder of markdown files. You read it in Obsidian, a free notes app that shows your files as a linked graph. Claude Code writes the files, Obsidian is how you look at them.

Step 1: clone the starter

Start from our starter, not a blank folder. It already has the CLAUDE.md schema, the folder structure, the skills, and the full guide inside. Copy this, paste it in your terminal:

git clone https://github.com/automation-flow/cba-starter business-brain
cd business-brain

See the starter on GitHub

Open that folder as an Obsidian vault, then open Claude Code inside it.

Step 2: tell Claude let's go

In Claude Code, run /onboard, or just type "let's go, set this up with me." From here you mostly answer questions. Claude:

• interviews you one question at a time, then writes your CLAUDE.md and the first wiki pages;
• offers to connect your tools, one at a time, when you're ready;
• turns on sync so the vault backs up and shares with your partner on its own.

Plan for about an hour. Most of it is you answering questions while Claude does the typing.

The most useful Claude Code skills

A skill is a saved instruction set that teaches Claude how to do one thing your way. Install it once, then trigger it with a slash command like /humanizer, or just ask in plain words and Claude picks the right one. These eight are the ones we reach for every day.

Two commands worth knowing

Skills aren't the only thing you trigger with a slash. These two Claude Code commands earn their keep from day one:

The Humanizer, up close

It's the skill we run on every word that leaves the building. Clone it, drop it in your skills folder, and Claude rewrites anything that smells like a machine wrote it. Give it feedback as you go, correct the lines that don't sound like you, and over a few weeks it settles into your actual voice.

mkdir -p .claude/skills && cd .claude/skills
git clone https://github.com/blader/humanizer

Meeting-prep: a client profile before every call

Give it a name and a company, get a full client profile instead of an hour of prep. Our rule: every claim has a source, and anything not public gets marked "to verify in the discovery call". Output comes back in Hebrew when the name is Hebrew.

How it works

1. Confirms on LinkedIn that the person actually works there now.
2. Researches the company live, news from the last 12 months, products and stack, with sources.
3. Reads the last 60 days of their posts for themes, tone, and ice breakers.
4. Assembles an 8-part profile into our Word template.

The profile template, 8 parts

We packaged the whole thing — the detailed agent and the Word template — as a public repo. Clone it, point your own agent at it, and you get the same 8-part profile shape we run before every call.

The agent in the repo is the production version with our internal plumbing removed — full research protocol, LinkedIn validation, and a strict no-fabrication audit. Clone it and run it with your own tools.

Don't forget about security

Once an agent can act on its own, send email, move money, message a client, the stakes change. Two habits keep you safe.

Instructions are not capabilities.

Telling an agent "never send email" is weaker than simply not handing it a send-email tool. Scope what it can touch. One team we know had an agent read a line of a task list as "send these," and it pushed promos to 150,000 inboxes nobody approved. It just could.

Keep your keys in a vault.

In practice: put every API key and token in 1Password, never in a file in the repo and never pasted into a chat. It has a CLI and secret references, so an agent uses a credential without the raw value ever sitting in your code or your git history. If a key ever leaks, rotate it.

The tools we use

Mapped to the three layers. Most of it has a free tier, start there.

Capture Record everything that happens.	Spinach AI: Transcribes every meeting automatically. Timeless: Captures every meeting and turns the key moments into action. Whisper Flow: Voice to text anywhere you type. Web Clipper: Save any web page to the vault in one click. gws CLI: Pull Drive, Gmail and Calendar into raw/.
Brain Turn raw data into linked knowledge.	Obsidian: Where you read and browse the wiki. Claude Code: Writes and maintains the wiki for you. Git / GitHub: Versions the vault and syncs it across the team.
Act Agents that read the Brain and do the work.	Claude Code: Runs the agents and skills. n8n: Automations and scheduled flows. NanoClaw: Open-source headless agents that run work in the background. Firecrawl: Gives agents reliable web access. Tavily: Search built for agents. Bright Data: Scraping and pages behind logins. Supabase: Postgres the agents read and write, via MCP. Vercel: Deploy the apps your agents build.

Going deeper, for the technical folks

# add a remote connector, then sign in
claude mcp add --transport http <name> <url>
/mcp        # finish the browser sign-in, see what's connected

./scripts/install-launchagents.sh