How We Learn New Skills Has Changed
AI certainly makes it easier to get moving with a new tool or learn something new.
You can describe what you’re trying to accomplish and get back something usable: a config, a snippet, a rough plan. With a little cleanup, it can save you hours.
Used well, that’s a legitimate advantage. Most of us aren’t learning in a lab. We’re learning in the middle of real work, with deadlines, constraints, and production risk.
The catch is that speed can hide what you didn’t actually learn. It’s possible to ship something that works and still be fuzzy on the mental model behind it. And that fuzziness shows up later when you need to debug, extend, or explain the approach to someone else.
You can use AI to both help you develop solutions quickly and learn new skills, but only if you frame it correctly. Treat it like a guide that helps you build understanding, not a shortcut that bypasses it.
The Three Layers of Learning
Learning a tool under pressure is mostly a sequencing problem: what to understand first, what to practice next, and what you can defer. Breaking it into three layers can help us identify where AI can help and where it should be left off.
1) Mental model (conceptual).
What is this tool trying to do? What problem does it solve? Without this, you can memorize the steps a tool needs to work and still feel shaky when something deviates from the happy path.
2) Workflow (procedural).
The repeatable sequence you’ll run most of the time. This is where examples and checklists are helpful, because they give you a default path to follow while you’re still building context.
3) Judgment (context + tradeoffs).
When to use it, when not to, what “good” looks like in your environment, and how you recover when things fail. This is the layer that turns “I can follow steps” into “I can operate this safely.”
AI is genuinely helpful on the first two layers. The place to be careful is the third. If the model is making the choices, you may move quickly, but you won’t build the confidence that comes from knowing why you chose that path.
If you want some internal framing on building “real ownership” over time, check out How to Become a Domain Expert.
Where AI Helps and Hurts Learning
Once you’re clear on the layers, you can use AI with more precision.
What AI is great for when you’re learning
- Translate and compress docs into your context. Ask it to explain the concept using your stack and constraints.
- Map the mental model. Have it list the key concepts you need before touching production, along with the common failure modes to watch for.
- Lay out the workflow. Get the 80% path plus the common deviations.
- Generate reps. Ask for small exercises and variations so you can practice more in less time.
This aligns with what we know about learning: durable learning comes from retrieval and re-application, not repeated exposure. The testing effect, as described by Roediger & Karpicke, is a classic example of retrieval beating re-study over time.
Where AI can create “shallow learning”
- Full solution before first attempt. You get output without building the mental model.
- Letting the model pick the approach. It feels like “best practice,” but you didn’t practice tradeoffs.
- Avoiding failure modes. You learn the happy path and get surprised later.
A useful concept here is “desirable difficulties.” Some friction is the point, because it forces the kind of effort that sticks.
There’s also a practical warning sign from the “Google effect” research: when we expect information to be available externally, we’re more likely to remember how to find it than the content itself. That isn’t automatically bad, but it explains why constant reliance can reduce what you carry in your head.
|
|
Where can AI save you time?
My friends at Big Creek Growth put together a quick survey to spot the repetitive work you can hand off to automation.
|
|
The Learning Ramp-Up Loop
If you want AI to help you learn (not just finish), you need a loop you can run on purpose. Something that keeps you moving while still forcing you to build the mental model and the judgment.
- Frame the job (you lead).
Define the outcome and constraints: scope, risk, environment, standards, and time. If you’re vague, AI will fill in the gaps with assumptions.
- Build the mental model (AI-assisted).
Ask for a plain-language explanation of what the tool is optimizing for, how the pieces fit, and what commonly goes wrong.
- Learn the core workflow (AI-assisted).
Get the 80% path and the “what people usually forget” list. If your team has conventions, include them. Generic guidance is less useful than “how we do it here.”
- Do one small rep (you own).
Build the smallest meaningful slice in your real environment. One endpoint. One flag. One pipeline step. One migration. Small enough to reason about end-to-end.
- Pressure test (AI challenges, you fix).
Ask for edge cases, misuse cases, and operational risks. What fails first? What breaks silently? What would you log? Then you add the tests, guards, and rollback plan.
- Prove ownership (you confirm).
Explain it back: how it works, why you chose it, what the tradeoffs are, and how you’d recover.
When AI is designed to guide learners (with questions, scaffolding, and feedback) rather than just dump answers, outcomes can improve. Harvard’s report on AI tutoring is a great starting point to learn where AI can help with learning and where it falls short.
If you’re learning a new tool or approach at work, it’s normal to want to get productive fast. AI makes that possible in a way we didn’t have 5 years ago.
To get value out of AI as a learning tool, you have to focus on building the mental model, the workflow, and the judgment around the tool or skill.
Pick one thing you’re ramping on right now and run the loop once. Keep it small. Build one slice. Pressure test it. Then try to explain it back without looking anything up.
If that last step is hard, it’s a signal you’re still early in the ramp, and the fastest way forward is another rep with the tool, not a more elaborate prompt.