Skip to content
Codrlabs Open

While you're learning

If you are here to get better — a first contribution, a mentored task, a new language or library — how you use AI decides whether you come out more capable, or just finished.

Friction is the point

Section titled “Friction is the point”

Expertise comes from doing the work: tracing an error with no useful log, rewriting an approach that will not scale, getting it wrong and finding out why. Lars Faye’s word for what that builds is Fingerspitzengefühl — fingertip feeling — the instinct that says “this is going to cause problems” before you can say why. Skip the struggle and the instinct never forms.

Anthropic’s researchers reached the same place from a randomised trial: “Cognitive effort—and even getting painfully stuck—is likely important for fostering mastery.”

What the studies saw

Section titled “What the studies saw”

Asking beats delegating. In Anthropic’s 2026 trial, 52 mostly junior developers learned an unfamiliar Python library, Trio. Those with AI scored 50% on the follow-up quiz against 67% for those coding by hand, with the largest gap on debugging. How people used AI mattered: those who asked conceptual questions and resolved errors themselves, or asked for explanations alongside code, scored well. Those who delegated the code, leaned on AI more as they went, or had it debug for them averaged under 40%.

Hints beat answers. Bastani et al. gave about 1,000 high-school students GPT-4 during maths practice. With a plain chat interface, practice scores rose 48% — but on the exam, with AI taken away, those students scored 17% below students who had only the textbook. A tutor version instructed to give teacher-designed hints rather than solutions lifted practice by 127%, with no significant harm on the exam. Students using the plain version did not perceive that they had learned less.

AI can widen the gap it should close. Prather et al. observed novices solving a programming problem with AI across 21 lab sessions. Students who were already progressing used it to write code they had already planned, and ignored bad suggestions. Struggling students had their existing difficulties compounded, thought they had performed better than they had, and “finished with an illusion of competence”.

How to use AI while you learn

Section titled “How to use AI while you learn”

Do:

  • Ask conceptual questions. “Why does this need to be awaited?” teaches. “Write the function” does not.
  • Ask for a hint, not the answer. Tell the assistant explicitly not to give you the solution.
  • Write it first, then ask for critique. Your attempt, then its review — never the reverse.
  • Use it as interactive documentation, then confirm what it said against the real documentation.
  • Explain your understanding back and ask what is wrong with it.
  • Read the error before you paste it. Form your own guess first, then check.

Avoid:

  • Generating the solution to the problem you are meant to be learning from.
  • Accepting code you could not rewrite without the assistant.
  • Letting the assistant fix errors it introduced, in a loop you no longer follow.

Before you prompt

Section titled “Before you prompt”

Adapted from the checklist Lars Faye uses in his own daily work:

  1. Without an AI tool, could I still do this task?
  2. Am I using the model to deepen my understanding, or to get to the answer faster?
  3. If I had to audit the output, could I explain what is happening?
  4. If this concept is new to me, have I read enough to know the right questions to ask?
  5. Have I checked the approach against documentation or another source?
  6. Is this a rote task done a hundred times before, or does it need a decision somewhere?

If the honest answers point toward getting the answer rather than understanding it, do it the slow way.

AI tutors are wrong too

Section titled “AI tutors are wrong too”

Using AI as a tutor carries the same risk as using it to write code: it can be confidently wrong, and a learner is the person least equipped to notice. As Faye puts it, if you cannot audit generated code, you cannot audit a generated concept either. Check what it teaches against official documentation, other people, and running the code yourself.

When you are stuck

Section titled “When you are stuck”

Being stuck is expected, and it is not wasted time. Give it a real attempt, then ask a person. In a mentored task, tell your mentor what you tried and where AI helped — that is the most useful thing they can know when they review your work.

In Stack Overflow’s 2025 survey, 75.3% of developers said they would still ask another person when they do not trust an AI’s answer. That instinct is right.

Sources

Section titled “Sources”