Two years working inside AI: The foundation

After two years of building systems with AI rather than around it, these are the principles I keep coming back to. Across workspaces, clients, and domains, the same patterns hold. Part 1 is the shift in how to think about the work. Parts 2 and 3 cover how the system runs and what to start this week.

1. People command. AI executes. Knowledge compounds.

The role of a person shifts from doing work to commanding work, defining what excellent looks like. Agents carry domain expertise, company knowledge, and a judgment framework tuned to the person who commands it. The agent does 90% of the work. Every correction lands in the knowledge file. Every future agent inherits it.

2. The model is the engine. You drive a car.

Claude, GPT, Gemini are roughly interchangeable. Picking one over another rarely changes the outcome. The system you build around the model produces the differentiation.

3. Your IP is your knowledge, not your tools.

Every competitor has the same models. Your edge is the knowledge you’ve captured: how your best people work, what good output looks like in your business, the rules you’ve encoded over time.

4. Hire agents the way you hire people.

Each agent needs four things: a job description, domain knowledge, rules of behavior, and the tools it can call. Build all four and you have something useful. Miss one and what you have is closer to a chatbot.

5. The bottleneck isn’t the model. It’s defining how your business runs.

The tooling is ready. What takes the time is defining how the job gets done, what good output looks like, and which frameworks to follow. Once defined, the system executes consistently.

6. Using AI makes you faster. Working inside AI changes everything.

Using AI keeps the jobs intact and speeds them up by some percentage. Inside AI swaps the jobs entirely. Agents take execution. People design the work and decide. The org chart gets redrawn.

7. Typing is over.

Dictation captures the qualifiers, the reasoning, and the “because” behind every decision. Typing strips most of that out. The simplest way I’ve found: schedule a Google Meet with yourself on a Saturday morning, walk for an hour, and talk through what good looks like in your business. An hour of unfiltered thinking transcribes into pages of IP that lives only in your head.

8. Your best performer’s brain is your most valuable asset.

Sit them down and press record. “Onboard me like I’m new.” An unfiltered conversation with your top performer captures more than every SOP your company has written. That transcript becomes the foundation for the agent that does the role.

9. One agent, one job.

A sales strategist. A pipeline analyst. An account manager. Each one focused on one thing. The agent asked to do everything produces average work.

10. Your agents live in Claude Code.

One workspace for all your agents, knowledge, rules, and workflows. Run them from Claude, Slack, or any chat.

If you got something out of this, share it with someone who would like to join the conversation. Follow Michael Chauliac and harperOS. Part 2 covers how the system runs: where knowledge lives, how agents read your data, and how to counter the failure modes.