Embrace The Ugly

Hire a new employee and tell me they never hallucinate. How often do you clean up after coworkers? How often do you have someone on the team who just doesn't care about the details? You accept human flaws every day. You can accept AI flaws too -- especially when you can actually fix them.

The best test I can give you: give the AI a task you already know how to do. Something you've been doing for 10 years. Compare the results. I bet you the agent does as good or better, if it's set up correctly.

You can also put guardrails around it. Tell it: if you claim something, prove it. Show me the link. Show me the source. Then you verify. You do that until you trust it in specific areas.

And here's the thing nobody wants to hear: when the AI gets it wrong, most of the time it's because you didn't tell it properly. We're terrible at giving clear instructions.

Here's a trick that fixes most problems: ramble at it. Say what you want in whatever messy way it comes out. Then ask: "Can you tell me what it is I'm asking you to do?" It says it back to you, structured and clear. You read it and realize that's not what you meant at all. So you correct it. You build the instructions together. This takes seconds.

First question I ask: where do you store your documents? Google Docs or OneDrive. Every time.

So you've already given all your data to Google or Microsoft. Every document, every spreadsheet, every email. But now you're worried about privacy? With AI?

If you trust Google, you trust OpenAI. If you trust Microsoft, you trust Anthropic. You're paying for the service. They say they're not using your data to train on. The privacy concern people have about AI specifically -- while their entire business lives on someone else's cloud -- doesn't hold up.

If you're bad at using a hammer, you're going to hit your finger. Same thing here.

If you say "write me a proposal," you'll get generic output. You gave it nothing to work with.

The fix: don't start writing proposals. Start by setting up a writing style. Give it 20 of your old proposals. Tell it: write like this. It will write exactly like you -- spelling habits, sentence structure, everything. You do this once. Then every proposal, every email, every document comes out in your voice. Forever.

People spend weeks comparing Claude Code vs Cursor vs Copilot vs Codex vs Gemini and while they're comparing, they're using none of them.

That's like walking 100 kilometers because you can't decide between a Ford, a Volvo, and a Tesla. Pick one. Any of them is so much better than doing everything manually.

I pick Claude Code. I think it's currently the best for business output. Anthropic's output is polished, business-ready, professional. Google's looks like a PhD student made it. For business, that matters.

You can argue with me. I don't care. It's my choice. If you like it, come with me. If you don't, pick another one. Just don't walk.

For almost a year, I've forced myself to go to Claude Code first for everything. I need to change the title of a Word document. I know how to open Word. It would take two minutes. I still go to Claude Code first.

In the beginning, your productivity will go down in some areas. Power through that. Because once you get past it, your productivity doesn't just double. I'm doing the job of a team of 10.

I'm a long-time freediver. My coach once said: don't jump in the ocean with a monofin and go spearfishing, because a shark will think you're an injured dolphin. A monofin swimmer is fast compared to a human. Compared to a dolphin? Still slow. People who don't use these tools are the monofin swimmers.

As a manager, you switch context maybe several times a day when employees come to you. Working with AI agents? You do that 200 times a day.

When I started, I could manage two agents at once. Now I run six to ten regularly. It's a muscle you build. I have maybe four to six good hours of heavy parallel work, then it drops off. At some point you just run out of good.

My garage has never been cleaner. When you're running six agents and waiting for responses, you do five-minute cleanups all day.

I've been programming since the late 1980s. Within a year of using AI agents, I no longer do what I trained to do my entire life.

My skill went from being a creator and writer of code to being a manager of six to ten occasionally drunk PhD students. There were genuine moments where I wondered: can I even call myself a developer anymore?

I played basketball my whole life at a high level. Knee injury ended it. Missed it for a long time. Started freediving instead -- completely different, meditative, quiet. Now I don't miss basketball at all. Things change. You become something different.

I haven't written a single line of code in probably six months. I don't miss it. I'm still a creator. I'm just a faster creator who doesn't go as low.