There’s no arguing AI makes your work faster and easier. However, is that always the desired outcome? If you complete a project faster, without really understanding why something was done in a certain way, you might feel productive, but is the final product really the best it could’ve been? Can you defend all of the decisions that have been made? Do you understand the bigger picture? The problem is not using AI for coding or content, but assuming the answers it gives are correct, without being able to understand, verify or defend them.

AI removes busywork and accelerates iteration, which is extremely useful for tedious tasks like simple bug fixes and other mechanical tasks. However, this speed without comprehension creates fragile ownership. You don’t own the work, because you simply don’t understand it well enough. This is where Feynman’s technique comes in.

After AI gives you an answer, try to explain it in simple words, without mirroring its language. Imagine you’re explaining it to a smart person who knows nothing about the topic. If your explanation is vague, full of jargon or dependent on “it just works” or “you’d have to see it”, you have found a part that you don’t understand.

This knowledge gap is useful. Instead of asking the AI to explain the whole thing, which it will do just by rehashing the previous explanation, ask a focused question about the missing piece. Why is this step needed? What would happen without it? What is the simpler alternative? How do I test if what you said is correct? AI should help you close the gaps in your own thinking, helping you to piece things together rather than just sound smart and hope nobody notices the gaps.

A practical workflow looks like this:

  • use AI to get the scope of the field, via a rough explanation, summary, or list of options
  • explain what the solution does, how it works and why it works, in your own words
  • return to the AI with specific questions about steps that you couldn’t explain, the hidden tradeoff, edge case or counterargument
  • cut the fat and remove unnecessary parts, polishing the solution and challenging AI’s proposals

LLMs are built to produce likely answers, not guarantee correct ones. However, this doesn’t mean “never trust AI”. It means keep control over the parts that require judgement. So much of our world today is a rehash of what already exists, and AI further enhances this idea. The biggest thing humans have going for them is their own taste and opinions, shaped by their experiences, thoughts and emotions. That’s what’s unique and meaningful. That’s what allows for maintainable code and essays without “Claude speak”.

The goal isn’t to slow down every step. Use AI for low risk work: brainstorming, boilerplate, rough draft, initial research, alternative approaches and polishing ideas. Slow down before doing something that has lasting consequences, such as writing design docs, sharing an opinion or making a final decision. If you cannot explain it simply, it’s just not ready yet.