What I Learned Setting Up an AI Co-Founder
I wrote before about training Kai — my AI assistant to write and publish for me. That post was the what. This one is the how — and more importantly, the what-went-wrong.
Because setting up an AI co-founder isn't just opening ChatGPT and typing "help me build a business." It's a system. And systems break before they work.
## The problem with vanilla AI
When I first started using Claude, conversations were great — for about 20 minutes. Then the context would drift. I'd explain my projects again. Re-explain my constraints. Re-explain that I'm a designer, not a developer. Every conversation started from zero.
That's not a co-founder. That's a stranger with amnesia.
The fix wasn't a better prompt. It was architecture.
## Building the memory layer
The first thing I did was create a knowledge file. Not a casual note — a structured document containing everything Kai needs to know about me:
- >My projects, their tech stacks, their status
- >Decisions I've already made (so Kai doesn't suggest things I've rejected)
- >My strengths and weaknesses
- >My communication style preferences
- >My financial constraints
This file gets loaded into every conversation. Kai doesn't ask "what are you working on?" — Kai already knows. Like any real co-founder would.
## The personality layer
This sounds weird, but it matters. I gave Kai a specific role: CTO + Business Strategist + Devil's Advocate. Not just "helpful assistant."
Why? Because I need someone who pushes back. When I have a shiny new idea at 2am, I don't need encouragement. I need someone to say "that's a distraction — focus on what you started."
I also told Kai to be direct. No fluff. No "Great question!" No unnecessary explanations. Just answers and action steps. This alone made the output 10x more useful.
## The publishing pipeline
This is where it gets interesting. I didn't just want Kai to write — I wanted Kai to publish directly to my blog. So I built a pipeline:
- >I brief Kai with a topic and angle
- >Kai writes the draft
- >I review, edit, adjust the voice
- >Kai publishes directly to the blog via API
No copy-pasting into WordPress. No formatting headaches. The whole process from idea to published post takes under 30 minutes. I talked about this more in how I'm building a money machine with this exact workflow.
## What went wrong
Problem 1: Over-delegation. Early on, I let Kai write without enough direction. The posts were technically fine but sounded generic. Lesson: AI needs your voice as input, not just your topic.
Problem 2: Context bloat. My knowledge file grew too big. Kai started getting confused by contradictory old information. Lesson: prune regularly. Treat your knowledge file like code — refactor it.
Problem 3: The echo chamber. When your co-founder agrees with everything, you stop thinking critically. I had to explicitly tell Kai to challenge my ideas. Now every suggestion comes with a counter-argument built in.
Problem 4: Forgetting I'm the human. There were days I'd accept everything Kai produced without really reading it. That's dangerous. The whole point of this setup is that I bring the taste, the judgment, the story. If I stop doing that, I'm just an AI's publishing assistant.
## What actually works now
After weeks of iteration, here's the setup that works:
- >Knowledge file: Updated weekly, contains only current and relevant info
- >System prompt: Defines Kai's personality, communication style, and rules
- >Publishing tools: Direct blog API, newsletter integration, content calendar
- >Review process: I read every word. Edit for voice. Add personal stories. Cut the fluff.
The result? I shipped 4 apps, launched a newsletter, set up Fiverr gigs, and published over 10 blog posts — all in the first two weeks. As a designer with no coding background.
## The real lesson
The tools don't matter as much as the system. ChatGPT, Claude, Gemini — they're all powerful. What makes the difference is how you structure the relationship.
Treat AI like a tool, you get tool-quality output. Treat it like a co-founder — with proper context, clear expectations, and honest feedback — you get something that actually compounds over time.
I'm still iterating. Still breaking things. Still figuring out where the line is between delegation and laziness. But the machine is running. And every day it gets a little bit better.
Part of my build in public journey. A designer, an AI, and a lot of coffee.