I write 0% of my code manually

I was chatting with friends recently. One mentioned that developers he knows had gone from writing ~30% of their code manually to ~5% in the last three or four months. He couldn’t confirm the numbers, but the direction was clear. I said that even 5% manual is underperforming. Whether that’s literally true or not, the bar for “good enough” with AI is moving fast.

If you want to get better at this, the usual advice is noise. Courses, tutorials, best practices, mostly theater. What actually forces improvement is simpler. Constraints.

Learning by Removing the Crutch

I learn fastest when I set strict limits on what I do and don’t do. From my junior days, fifteen years ago. I didn’t know Vim. To get good, I turned off the mouse and disabled the arrow keys. The first few weeks were awkward, but those limits forced me to learn properly. No fallback. I got good at Vim because I removed the easy escape hatch.

That’s a forcing function. You don’t “practice when you remember.” You remove the old way. The only path forward is the new one.

Applying It to AI

I do the same today. My rule is to write 0% of code manually. No implementations, no micro-edits. Nothing. I’ve done this since October. Initially slower, but the limit pushes me daily to learn and get better. I call it becoming an “AI cowboy.” Corny, but the mechanism is real.

If you can always fall back to typing, you never learn to steer the AI. You never improve prompts, context, or workflow. You just type the fix and move on. Zero manual code forces you to fix the real bottleneck. How you work with the machine.

What to Expect

Start by expecting a calibration period. You’ll know exactly what to type while the AI misses details. Instead of taking over, tighten the prompt, add better context, and iterate once more. That’s where the skill compounds. Over time, “I could have typed this faster” shrinks. “The AI got it, I just reviewed” grows.

Yes, you do need literal 0%. No half measures. Pick the limit, hold the line.

If you want to get better at AI, try it. Pick a constraint that feels slightly too strict. Give it a few weeks. See what breaks and what you learn.