20.02
item · Curriculum

Learning with AI: a working list

Most developers use AI to get answers. Fewer use it to get understanding. The difference matters. An answer solves one problem. Understanding changes how you approach every problem after it.

What follows are concrete techniques — things you can do today, with whatever AI tool you have access to. They’re organized roughly from “I’m stuck” to “I want to get better.”

Building a mental model

Explanation on demand

“Explain what this code does.” Simple enough. But the real power is in the follow-up: “That didn’t click — explain it a different way.” Then: “Give me an analogy.” Then: “What would happen if I changed this part?”

AI has infinite patience for reframing. A human mentor might explain something three ways before getting frustrated. AI will explain it thirty ways without breaking a sweat. Use that.

Mental model verification

This one is underused:

“I think React re-renders a component whenever its parent re-renders, regardless of whether the props changed. Is that right? Where does my mental model break down?”

Most developers either assume they understand something or don’t examine the assumption at all. Explicitly asking an AI to test your mental model surfaces the gaps you don’t know you have.

Ask for the questions

Sometimes you read code, nothing in it is unfamiliar, and you still don’t get it. You can’t ask a good question, because you don’t know what you’re missing. So ask for the questions:

“I’m comfortable with RxJS but new to signals. What questions should I be asking about this component?”

This works better than “review this code.” A review is about the code: what’s wrong, what could be better. Asking for questions is about you: the gap between what the code does and what you understand about it. The questions tend to land where it matters. Why is this a computed and not an effect? What happens if this runs twice? Who else writes to this state?

Say who’s asking. The questions get sharper when the AI knows what you already know.

And write your own first. Before you ask, jot down two or three questions of your own, then compare. The ones you didn’t think of are the lesson. Coming up with good questions is a skill, and it only grows if you practice it.

Narrate the runtime

“Walk me through what happens, step by step, when a request hits this endpoint. What runs, in what order, and what’s in memory at each step?”

Code is written in one order and runs in another. Middleware, dependency injection, change detection, the event loop, a handler that picks up a message on another thread: the hard part of a framework is usually knowing what runs when. A narrated trace turns the file you’re reading into the sequence that actually happens. Then check it in the debugger. A trace is a claim, not a recording.

The “why” behind the “what”

“Why does Angular use dependency injection instead of just importing services directly? What problem was it solving?”

Understanding why a design exists is more durable than understanding what it does. The “what” changes with every version. The “why” usually doesn’t.

Comparative explanation

“Show me three ways to handle form validation in this component and explain the tradeoffs of each.”

This forces exposure to a design space rather than a single answer. You don’t just get a solution — you get the judgment to choose between solutions. That judgment is the skill that survives when the APIs change.

Learning from failure

Deliberate breaking

Write or generate something that works, then ask:

“What are the ways this could fail? Show me an example of each.”

This is particularly valuable for frameworks where you can get surprisingly far without understanding the underlying model — until something breaks in a confusing way. The developer who has seen the failure modes, even artificially, is the one who can debug them under pressure.

Failure mode study

When something breaks, resist the urge to ask the AI to just fix it. Instead:

“Explain why this broke. What mental model would have predicted this failure before it happened?”

The fix is cheap. The model is the asset. A developer who understands why something failed can prevent the next ten failures. A developer who only knows what to change will fix them one at a time.

“What am I not seeing?”

Show the AI working code and ask:

“What subtle problems or edge cases could bite me later that aren’t obvious right now?”

This surfaces the category of issues that normally only experience reveals. You’re borrowing decades of collective pattern-matching from the AI’s training data. It’s not infallible — but it’s a remarkably good first pass at the question “what will future-me regret about this?”

Going deeper

Close the cargo-cult gaps

Every developer has things they use daily without really understanding:

“I’ve been using useEffect for months but I think I’m cargo-culting it. Help me actually understand what it does and when I should and shouldn’t use it.”

AI makes it easy to close these gaps on demand. The cost of admitting “I don’t really understand this thing I use every day” is zero when your only audience is an AI. Take advantage of that.

Concept mapping

“I just learned about observables. What other concepts are closely related that I should understand to use observables well?”

This builds connected knowledge rather than isolated facts. It’s the difference between knowing a word and understanding the language.

Active recall

“Explain closures to me, then ask me some questions to test whether I actually understood.”

Passive reading feels like learning. Active recall actually is. AI can play the role of a tutor who explains, then quizzes, then explains what you got wrong and why — with no judgment and no time pressure.

Personalized learning

Custom learning plans

Share your code or describe your work, then ask:

“Based on what you see here, what concepts, patterns, or language features would most improve the quality of what I’m building? Give me a prioritized learning guide.”

The results can be surprisingly tailored. AI can look at your actual code and identify specific gaps — not generic “learn these 10 things” advice, but “you’re doing X manually in three places; if you understood Y, all three would simplify.”

Pre-class preparation

Before a training session, workshop, or even reading documentation:

“I’m about to learn about message queues. What’s the core concept, what’s the hardest part to understand, and what prior knowledge should I make sure I have?”

This is the equivalent of reading the chapter headings before the lecture. It gives your brain a scaffold to hang new information on.

Scaffolded practice

Write first, then compare

Attempt a solution without AI help, then ask the AI to solve the same problem and explain any differences.

The comparison is often more instructive than either approach alone. You see not just what’s different, but why — and you can interrogate the differences: “You used a map here where I used a for loop. Is there a meaningful reason to prefer one?”

AI as reviewer, not author

Write the code yourself, then ask the AI to review it:

“Review this code and explain what could be better and why.”

This keeps you in the production role while getting feedback that would otherwise require a senior colleague’s time. The AI won’t always be right, but it’s always available, and the act of evaluating its suggestions is itself a learning exercise.

Challenge its choices

When AI solves something differently than you would, resist the instinct to simply override it. Instead:

“I tend to do this differently — are there advantages to your approach, or disadvantages to mine, that I should consider?”

This works because it’s organic — triggered by real code, not a hypothetical exercise. The result is either a genuine update to your practice, or increased confidence that your preference is well-founded. Either outcome makes you better.

Note. The value of writing code by hand — even when AI can generate it — is probably not about the output. It’s about building the mental models, and especially the failure intuitions, that make you capable of evaluating AI-generated code. How much hand-writing is needed to build a sufficient model is a question worth revisiting as the tools evolve.

Exploring the problem space

Rapid iteration

AI is well-suited to the “finding out” phase of development — quickly generating alternatives, prototypes, and variations so you can evaluate options before committing to one.

The criticism that “AI doesn’t write maintainable code” often misses the point: the expensive mistake isn’t writing messy code, it’s writing clean code for the wrong thing. Optimization for maintainability belongs to the phase after you know what you’re building. AI accelerates getting there.

A/B testing approaches

“Generate two implementations of this with different tradeoffs — one optimized for readability, one for performance — and explain the differences.”

This is faster with AI than any other approach and builds real judgment about design decisions rather than defaulting to habit.

One underappreciated gift AI offers is making code less precious. When generation is cheap, the real work — understanding the problem, making the right tradeoffs, knowing when something is good enough — moves back to the center where it always belonged.

Keeping it yours

Close the chat and explain it

Every so often, close the conversation and explain what you just learned: out loud, to a colleague, to a rubber duck, or in a commit message. If you can’t, you haven’t learned it yet. You’ve borrowed it.

AI doesn’t make anyone worse at this work by default. It does make it very easy to borrow understanding instead of building it, and borrowed understanding is gone the moment the chat closes. The test is cheap, and it’s the one that matters.


These ideas were developed in conversation with Claude. It would be a strange kind of dishonesty to write about AI-assisted thinking while hiding that I used it.