What I learn now that AI writes my code

In 2025 I used AI to learn to program, and I learned Python, React and how to design a whole system. In 2026 I use it to build, and in a month I haven’t learned one new thing about programming. I don’t even know what’s in my own code anymore.

Both years were the same project, LLM Monster Hunter, my game where an AI makes up every monster as you play. In 2025 I built it with Claude Sonnet 4 in a chat window. In 2026 agents write it. I use Claude’s Fable 5 model inside Claude Code, Anthropic’s agent that works on your computer by itself.

So whether AI made me learn came down to what I aimed it at. What I learn now is how to run the tools, and I’d make that trade again. I’m just not sure I’d still notice if the model were wrecking something.

What I learn now instead

• Which tool for which job. Claude Code, Cowork (Claude’s desktop agent for work that isn’t code) or the plain chat. They connect to different things.
• Which model, at which thinking level (how long it thinks before it answers), and what that combination will really produce.
• What can run without me, and which parts still need me to make the call.
• Techniques. Skills (saved instructions an agent loads for one kind of job), workflows that run a team of agents, and lately the gauntlet loop, where a fresh critic agent judges each piece blind.
• Orchestration. OpenClaw, an always-on personal agent, runs on a schedule and does web crawling and standing tasks for me.

Lessons from the last month

• Point the model at your docs instead of pasting everything in. Keep each thing in the right file and tell it where to look, so you don’t pay for all of it on every message.
• Use far fewer rules than you think. It already knows most of what I used to spell out, and a rule written in advance makes it worse at the case you didn’t see coming.
• But say what you both take for granted. Otherwise every agent on a long run spends real thinking working it out again.
• Be clear about what you expect. That only works if you know what the model can do.
• On a long job, ask for the presentation at the start. Then it takes screenshots and notes as it goes. That’s how my agent handed me a narrated recap of my game.
• Work in bursts. Send off hours of work, walk away, and review it the next day. More on that in why I stopped checking on my agents every hour.

Aimed at learning, AI taught me to program

People say AI makes you dumber or a worse programmer, and that’s not what happened to me. I couldn’t program. I’d wanted to since I was a kid and failed at it for fifteen years. Then I used AI to learn, and I did. I learned a full stack (every layer of an app, front to back) and designed an architecture I can still defend. For a summer I read every line that went into my game, and I held the whole thing in my head because nothing else was holding it. It was hard and slow and I could feel my brain working. Later I aimed it at building more, faster, and got exactly that.

Why I’d make the trade again

What I learned in 2025 is why 2026 worked. A model can only run far ahead of you on a foundation that’s already clean and documented, and I’m the one who made mine clean. I think that’s why the $200 week that restarted my game worked.

What I don’t know is whether that stays true. If I never build another foundation by hand, will I still know enough to tell when the model is wrecking one? A month in, I don’t have that answer. For now I’m choosing to know the tools inside out and let the model do everything it can do for me.