Your AI coding tool finishes the feature. The button works. You can finally move the task out of the pile.
Then someone asks why it works that way, and you realize the explanation left with the chat window.
That is the beginner problem I kept thinking about while reading IBM's new workforce skills announcement, published September 21. In surveys conducted with Oxford Economics from April through June, IBM gathered responses from 1,500 human-resources leaders and 8,800 full-time employees. Sixty percent of employees worried that AI was eroding their skills.
That is a report of concern, not a controlled measurement proving that AI caused people to lose ability. The distinction matters. A scary percentage should not become a reason to abandon a useful tool.
For me, it raises a practical question: when AI helps you finish work, what part of the work are you still learning?
I came into software through a software engineering master's after assembling the prerequisite foundation. I also use AI to build apps. I want beginners to have that help. I want you to gain enough understanding to recognize when the help has gone sideways, too.
Here is the rule I would use: give the feature and your understanding separate finish lines.
“The list sorts correctly” is a feature result. “I can explain the sorting rule and predict what happens when two values tie” is a learning result. You can have the first without the second.
Before your next build session, choose one thing you want to understand by the end. Keep it as small as “why this list changes order.” If turning a broad app idea into a manageable first task is the obstacle, my $1 AI App Builder Starter Prompts offer a guided starting point. Add your learning finish line alongside the task.
What the research can actually tell you
There is a more direct coding experiment behind this concern. In research published in January, Anthropic randomly assigned 52 developers learning an unfamiliar Python library to work with or without an AI assistant. The AI group averaged 50% on a subsequent quiz; the other group averaged 67%. That is a 17-percentage-point gap. The completion-time difference was not statistically significant.
This was a small study of a particular learning task, with a quiz shortly afterward. It does not establish the long-term effect of every coding assistant on every developer. Its analysis of different interaction styles was observational, so those patterns do not prove that a particular prompting technique caused better learning.
A separate Microsoft Research study presented at CHI 2025 surveyed 319 knowledge workers. Higher confidence in AI was associated with less reported critical thinking. Again, these were self-reports and associations, not proof that using AI inevitably makes you worse at your job.
My response is a work habit you can try, not a scientifically validated recipe: keep a small part of each unfamiliar task available for practice, then check what you can do with it yourself.
Pick one mechanism, not an entire subject
Imagine you are building a fictional freelance task tracker. You ask AI to sort unfinished tasks by due date.
“Learn programming while building this” is too broad to check. I would choose a narrower target:
“I can explain how this app orders tasks, including a missing due date and two matching dates.”
Write down the intended behavior first. For this example, suppose tasks with dates come first, earlier dates come before later ones, and matching dates use the task title as a tie-breaker. Tasks without dates go last. That is a product choice for the example, not a universal sorting rule.
You now have something small enough to reason about. You do not need to understand the entire framework, deployment system, and every file AI touched in the same afternoon.
Predict before you see the answer
Make three pretend tasks. Give one an early date, one a later date, and one no date. Write the order you expect before asking the assistant to show the result.
Then try two tasks with the same date. Explain what should break the tie.
If you cannot predict the result, name the missing concept. Maybe you do not understand how two values are compared. Maybe you never decided what an empty date means. Those are different questions, and now you can ask the useful one.
A prompt I would use is:
“Help me understand this sorting rule. Ask me to predict the order for a small example before showing the answer. If I get it wrong, point out the first mistaken assumption and give me a simpler case. Keep the exercise separate from production data.”
The value is the prediction you commit to. An explanation can feel obvious once it is sitting in front of you.
Explain the mechanism after assistance
Let AI help implement the feature. Then put the generated explanation aside and write your own short account of what happens.
For the task tracker, that might be:
“The comparison handles missing dates first. When both dates exist, it compares them. Equal dates fall through to a title comparison.”
Point to the relevant code or logic as you explain it. If your account says missing dates go last but the code never handles them, you have found a gap worth investigating.
I would ask AI to critique that explanation against the implementation and documentation. I would also run the tiny examples myself. An assistant agreeing with you is useful feedback; the actual result is another necessary check.
Keep the explanation short enough that you cannot hide behind a paragraph of terminology. You are trying to understand a mechanism, not audition for a glossary.
Change the example and remove the answer
Now try a variation without the previous solution visible.
Suppose the product owner wants the latest due date first, while tasks with no date still belong at the bottom. Predict what changes and what must stay the same. Try the small adjustment in a local practice copy, then run your examples.
Simply reversing the finished list would move the undated tasks to the top. That is why a changed example tells you more than reciting the original answer.
If you get stuck, ask for a hint. After resolving the confusion, attempt another small variation yourself. At your next work session, revisit the idea briefly before reopening the explanation.
I would keep a tiny note: the concept, what I predicted incorrectly, and the next variation to try. “Still need help with missing values” is a useful result. It tells you where the next practice belongs.
Keep the cost small and deliberate
This approach takes time. You will sometimes finish a task more slowly because you kept one piece of reasoning for yourself.
I would reserve that effort for concepts I expect to reuse or need to review responsibly. Familiar formatting work can receive more automation. An unfamiliar decision that shapes the app deserves a closer look.
During an urgent delivery, you may need experienced review and an assisted fix first. Record the gap and return to it in a safe practice setting. Do not turn a live client incident into an improvised classroom.
Being able to explain a mechanism also does not establish that the whole app is ready to ship. Normal testing, review, and release checks still have their own job.
For your next session, choose one small concept. Predict an example. Explain what the assisted work actually does. Then try a changed example with the answer closed.
A useful result is one you can ship responsibly. A useful learning session leaves you better equipped for the next result.
Put the lesson into your next build
If you want an immediate guided action, use the $1 AI App Builder Starter Prompts to define one small build task, then add the learning check from this article.
For the organized path from idea to publication, AI App Builder From Zero is the $9 e-book. It covers planning, scope, architecture, prompting, QA, and launch, and includes all 40 Starter Prompts as a free bonus inside its PDF and EPUB. You do not need to buy the standalone prompts as well.
Review Radar is coming soon: view the preview and join the email waitlist. It is being built to bring research, tailored app screens, and an AI-ready project folder together around your idea. That can give your build more concrete material to work from; it does not replace your understanding, testing, or product decisions. Checkout is not open.
You can also find me here:
Medium: https://medium.com/@marcusykim
DEV.to: https://dev.to/marcusykim
Website: https://marcusykim.com/
X: https://x.com/marcusykim
LinkedIn: https://www.linkedin.com/in/marcusykim/
Upwork: https://www.upwork.com/freelancers/marcusykim
Top comments (0)