nick taylor / may 2026

Getting Started
with AI Without
Losing the Fundamentals

How to use AI as a learning tool without losing the fundamentals.

01 / 13

whoami

Nick Taylor

Developer Advocate at Pomerium
GitHub Star, Microsoft MVP, AWS Community Builder
20+ years in tech, 10+ years open source
From Montreal, Quebec, Canada
nickyt.co · @nickytonline everywhere

02 / 13

audience.poll()

Let's see who's in the room

Q1 Who's actively learning fundamentals right now? HTML, JS, data structures, etc.

Q2 Who's already using AI tools in their learning or day-to-day work?

Q3 Who's trying to do both at the same time?

Q4 Is cost a barrier to using AI tools for you?

03 / 13

identify.tension()

juniors

Told to use AI before building mental models
No baseline to know when AI is wrong
Pressure to ship fast // but at what cost?

seniors

Already have the mental models
Can usually tell when AI output is wrong
Even they are still figuring it out

AI is a force multiplier. Like the super soldier serum in Captain America, it amplifies what's already there. Good fundamentals get amplified. Bad habits get amplified too.

// Matt Pocock at AI Engineer: "Bad code is more expensive than ever - if your codebase is a mess, AI makes it worse faster."

04 / 13

fundamentals.status()

Your brain still needs to hurt

The struggle is the learning. Sitting with a problem until it clicks - that's how intuition gets built. AI can't shortcut that process. It can support it.

This isn't anti-AI. It's about using AI in a way that builds your skills instead of bypassing them.

Make mistakes. Debug them. Understand why.
Use AI to explain, not just to generate.
You need to be the quality filter it doesn't have.

05 / 13

warning.emit()

Copy-pasting AI is Stack Overflow
on steroids

stack overflow

Upvotes signalled quality
Comments called out bad answers
Dates showed age of the answer
You at least had to read it

AI

Same confident tone whether it's right or wrong
No upvotes, no comments, no dates
Trained on a lot of bad code
Output looks clean so you trust it more

// The feedback loop that would have taught you something never fires.

06 / 13

ai.setMode("tutor")

Use it like a tutor, not an answer machine

            // ✗ you learned nothing

            "Write me a function that does X"

            // ✓ this is learning

            "Explain why this code works"

            "What's wrong with my code?"

            "Explain this like I'm new to programming"

            "What would happen if I changed this part?"

Challenge it. Ask for references, links, citations. AI can be very confident and very wrong. "How do I know this is right?" is a skill that matters for your whole career.

07 / 13

develop.curiosity()

Get good at asking questions

Vague question gets a vague answer. This isn't new. Better questions get better answers whether you're asking a colleague, searching Stack Overflow, or prompting an AI. It's a skill that gets better with experience.

Read the code it gives you. Always. If you can't explain it to someone else, ask it to explain it to you.
Prototyping is genuinely easier now. Want to see if something is possible? Use AI freely here. The bar for "is this worth trying?" has dropped.
Just trying an idea? Fine to not read it line by line. // But know the difference.

// Used to be "Google it." Now I lean on an LLM. Both are valid.

08 / 13

llm.explain("context_window")

Context windows: short-term memory

The context window is the total number of tokens a model can process at once. A token is roughly ¾ of a word. Everything counts: system prompt, instructions, conversation history, files, images, and the response being generated.

The model can only work with what's inside that window. Anything outside it doesn't exist to the model.
1 million token context ≈ 750,000 words ≈ a ~750-page novel.
Dumping everything in isn't always the answer. Being deliberate matters.

09 / 13

ai.configure("skills_and_prompts")

Skills: configurable behaviour

You can give AI persistent instructions that shape how it behaves: its tone, what it focuses on, what to avoid. This is how you treat it less like a chat box and more like a configurable tool.

//

System Prompts

Instructions the model reads before the conversation starts.

{ }

Custom Instructions

Preferences for tone, format, and focus. Set it once.

→

Skills

Reusable context you give the model about your codebase or workflow.

10 / 13

import { mcp } from "model-context-protocol"

MCPs: AI beyond the chat box

Model Context Protocol lets AI connect to tools and services - your filesystem, APIs, databases, dev tools. It's not just a chat box anymore.

~~

Tool Calling

Model decides to call a tool. Harness runs it. Result goes back into context.

{}

Schema Only in Context

The model never sees your function code. Just name, description, params.

// Just plant the seed. You'll get here.

11 / 13

tools.resources

Resources

Ollama Run models locally. Free, private. * Needs a decent machine. Quality gap vs frontier models is real.
NVIDIA Build Free API access to frontier models. Good starting point.
OpenCode $5 first month, $10/mo after. Solid starter plan for AI-assisted coding.
CommandCode Go or Pro plans. Worth checking out.
Claude Code Use with local models via Ollama. No subscription needed.
AI Hero Matt Pocock's site. LLM Fundamentals, MCP tutorial, AI Engineer Roadmap, 75+ posts. Mostly free.

12 / 13

session.close()

Wrapping Up

You still have to do the work. Brain needs to hurt.
AI is most useful when you understand what it's doing wrong.
Use it like a tutor. Ask it to explain, challenge it, read what it gives you.
Cost isn't a barrier. Free options exist and are good enough to start.

@

One Tip a Week

one-tip-a-week.beehiiv.com
This week: use Claude Code for free

nickyt.co · @nickytonline

13 / 13

Getting Startedwith AI WithoutLosing the Fundamentals