The four C's: building a second brain with Claude Code

I keep coming back to one idea: the model isn't the product. The system is.

Watching Nate Herk walk through his "AI operating system" recently crystallised something I've felt building agentic platforms at work — the leverage isn't in which model you run this week, it's in the second brain you build underneath it: the folders, files, skills and routing logic that any coding agent can read. Swap Claude for Codex tomorrow and it all still works, because what you've really built is your own IP — your business and your way of working, written down.

Here's the framework that stuck with me. Four C's, in order. The first two are the second brain; the last two are the operating system.

1. Context — who you are. Your CLAUDE.md isn't a personality file, it's a router. It points the agent to the rules, references, skills and projects. The pulse check is simple: if you asked your agent something about your business, would it answer like a stranger or like a co-founder? If it's the former, your context is thin.

2. Connections — live data. Static context (background, past work) gets you part way. The rest is data that's always changing — your CRM, your inbox, your P&L. Wire those in through APIs and CLIs, not screenshots. I lean on APIs/CLIs over MCP servers for the same reason he does: more control, usually cheaper.

3. Capabilities — skills and agents. This is where it becomes an OS. A skill doesn't have to be a ten-step workflow; it can just be a prompt you keep reusing. The mindset that matters: a skill is never finished. Every time you run one, that's data — "I liked this, not that, update the skill." Treat one agent like one station on an assembly line, chain the outputs, and let cheaper models (Sonnet, Haiku) handle the parallel grunt work.

4. Cadence — running while you sleep. Automation is the last C, and you have to earn it. More autonomy means more cost, more risk and more maintenance. Which brings me to the part every marketer should tattoo somewhere: a prompt is never a permission layer. If an agent can send the email, assume it will. The cautionary tale in the video — an agent that picked up a task wrong and blasted a discount code to a list of ~150k people — is exactly the kind of thing CRM teams should lose sleep over. The fix isn't a sterner prompt. It's keys, not prompts: scoped access so the agent simply can't open the door to the room it shouldn't be in.

A few habits that make all of it compound:

• Have it interview you. A "grill me" skill that asks 20–30 questions pulls knowledge out of your head and into the system faster than you'll write it down voluntarily.
• Make it verify its own work. Don't accept the 70%-there first pass. Give the agent a way to check itself — open a browser, click through, test the personas — and you get 90%-there instead.
• Treat every slip as data, not failure. Fix the instruction once; it never happens again.

For me the MarTech angle writes itself. Everything I build — campaign agents, QA automation, lead workflows — lives or dies on these four C's. The teams that win the next couple of years won't be the ones chasing every model launch. They'll be the ones who quietly built a second brain, earned the trust to automate it, and put keys where the prompts used to be.

Inspired by Nate Herk's walkthrough of his Claude Code AI operating system.