Headwork

Guides

Stop relying on AI for coding: keep your skills while you use it

Do not ban the tools. Change what you ask them. Make your own guess at the cause or the design first, ask the AI for the idea before the code, and never accept a change you cannot explain in one line. That is the pattern that kept people learning in the one trial that looked.

What the trials found

The routine

  1. Guess before you ask

    Write one line: what do I think is wrong, or how would I build this? Thirty seconds. It turns the AI's answer into feedback on your thinking.

  2. Ask for the idea, then the code

    "Explain the approach and what could go wrong. No code yet." Read it. Try to write it yourself. Ask for code only for the part you are stuck on.

  3. Predict before you run

    Before running anything the AI wrote, say what it will do. When you are wrong, you have found the thing worth learning.

  4. One line per accepted change

    For every change you keep, write one line on why it works: in the commit message, a comment or your notes. If you cannot, you do not own that code yet.

Pick "Coding" in the coach setup. It tells the AI to ask what you have tried before it gives code.

Small rules developers already use

Questions

Is it bad to use AI for boilerplate?

No. Handing off code you already know how to write costs you little. The risk is in handing off the parts you have not learned yet.

I am a junior. Should I avoid AI completely?

You do not have to. Use it as the senior who asks "what have you tried?" rather than the one who takes the keyboard. Conceptual questions are where people in the trial kept learning.

How do I know if my skills are slipping?

Try a small task with no assistant once a week. If it feels much harder than it did a few months ago, shift more of your asks from "write this" to "explain this".

Next: limits that work, and where blocking fails