← Back to posts
Happy Friday Friends!
Here's what's inside:
Why AI-first learning is quietly breaking the next generation of developers, the ladder I'd follow if I were starting from zero today, and the honest reason I still code without AI on some of my client work.
Yesterday I opened a client’s codebase and spent two hours reading through it before I touched anything.
I didn’t paste it into Claude. I didn’t ask for an explanation. I read it. Top to bottom. Figured out where the state was living, where the routes connected, and where the auth flow broke down. Then I made the changes by hand.
I ship with AI. I used AI with my iOS app with Claude Code. I’ve used Cursor. I pay for Claude Pro. I am not about to tell you AI is bad or that real developers write everything from scratch. That post is not the post I’m writing.
I’m writing this because I keep seeing people in my DMs and my Discord who are learning to code through AI from day one, and I’m worried about what happens to them in two years.
If that’s you, read the whole thing. Nobody else is going to tell you this, because everyone selling you a course, a tool, or a tutorial right now has a financial reason to sell you the opposite story.
The trap
Open any “how to learn coding in 2026” guide and you’ll see the same pitch. AI makes learning faster. AI is a free personal tutor. Bootcamps cost twenty thousand dollars and AI costs zero. You don’t need to grind fundamentals anymore, you just need to build stuff, and AI will fill in the gaps.
Some of that is true. None of it is the whole truth.
Here’s what actually happens when you learn to code with AI doing the heavy lifting from day one.
You feel fast. You ship projects. Your GitHub looks busy. You post the screenshots. You feel like a developer.
Then you hit the first real wall. Your deployed app breaks in production. Your authentication flow has a race condition. Your database query is returning the wrong user’s data. You paste the error into Claude. Claude gives you a fix. The fix doesn’t work. You paste it back. Claude gives you a different fix. That one breaks something else. After forty minutes you realize you don’t actually understand what your code is doing, and neither does the AI, and you have no way to tell which of you is wrong.
That’s the trap. Not that you used AI. That you built on top of it before you had anything underneath.
The people selling you AI-first learning are not going to warn you about this moment because it comes six months later, after you’ve already signed up. They have no incentive to tell you that the projects you’re shipping right now are teaching you almost nothing about how code actually works.
What code actually is
Code is not syntax. Syntax is the easy part. Claude can write syntax in any language on earth.
Code is thinking made explicit. It’s you taking a fuzzy problem, breaking it into pieces small enough to be unambiguous, and ordering those pieces so a computer can execute them. That is a skill. It is the only skill that matters. And it is the skill that does not transfer when you let an AI do it for you.
You can watch a thousand hours of someone else lifting weights and you will not get stronger. You can read a thousand pages about how to ride a bike and you will fall over the first time you get on one. Coding is the same. The learning happens in the muscle. The muscle is built by doing it yourself, badly, over and over, until it starts working.
AI is a tool for developers who have already built that muscle. It is a terrible tool for building it.
The brutal part about the job market
I know what you’re thinking. The job market is cooked. Juniors can’t get hired. Why does any of this matter if there aren’t any jobs anyway.
Entry-level coding jobs are shrinking. Senior roles are growing. Companies are cutting juniors and keeping seniors because a senior engineer plus Claude Code can do the work of three juniors for a fraction of the cost. That’s real and it’s not going away.
When companies do hire juniors in 2026, they’re hiring for a very specific thing. They want someone who can sit down with a codebase they’ve never seen, figure out what’s happening, and ship a working change without supervision. That is a hand-on-keyboard skill. It is tested in interviews. You cannot fake it. And the people who learned to code by prompting cannot do it.
The juniors who get hired right now are the ones who can code. Not prompt. Code. The pool is smaller, and the bar is higher, and the only way through it is to actually build the skill.
If you’re starting from zero in 2026, the opportunity is not to get there faster. It’s to get there better than the thousand other beginners who are speedrunning tutorial projects with Cursor and calling it a portfolio.
The ladder
Here is what I would do if I were starting from zero today. This is not what I did, because when I started, AI tools were not this good. But this is what I would do now, and it is what I am recommending to everyone who asks me.
Days 1 to 30. No AI in the loop. At all.
Not Google. Not Stack Overflow. No documentation. Just no AI. No ChatGPT explaining concepts to you. No Copilot autocompleting your functions. No, Claude summarizing an error message. Turn the extensions off. Sign out of the tabs.
Pick one language. JavaScript or Python. Write your first hundred programs by hand. They will be bad. You will spend twenty minutes on errors that Claude would fix in four seconds. That is the entire point.
What you are building in these thirty days is not a portfolio. You are building the part of your brain that reads code and understands what it’s doing. That part of your brain only grows under resistance. Take the resistance away and it does not grow.
Days 30 to 90. AI as tutor only. Not as writer.
Now you can bring AI in, but under strict rules. You can ask Claude to explain a concept you’re stuck on. You cannot ask it to produce code for you. You write every line. If you don’t understand what a line does, don’t write it.
This is the phase where most beginners would rather die than continue. The AI is sitting right there. It could write the whole thing. You are deliberately handicapping yourself. It feels stupid.
It is not stupid. You are practicing the thing that will make you employable. The ability to write code yourself, with understanding, while an AI sits next to you and could theoretically do it faster. That is the skill companies are hiring for. That is the skill you cannot build any other way.
Days 90 to 180. Autocomplete earned.
Now you can turn on Copilot, Cursor tab-complete, whatever. The AI can finish your lines. It cannot start them. You write the function signature, you write the first line, you think about what you want the next line to do. The AI suggests it. You evaluate. You accept or reject.
You are still doing the thinking. The AI is typing faster. That is the correct relationship.
After six months. Agents, with eyes open.
Now you can bring in Claude Code, Cursor Composer, all of it. By now you can read generated code and tell when it’s wrong. You’ve spent six months building the muscle that lets you evaluate what an AI produces. You have earned the leverage.
You will notice you are better at using these tools than the person who started with them on day one. Dramatically better. You catch hallucinations they miss. You review PRs that the AI generated and spot bugs they would have merged. You know when to let the agent run and when to take the keyboard back. The skill you built by coding without AI is exactly the skill that makes you good at coding with AI.
Why I still do this on client work
I am not on day 180. I have been coding for years, I am finishing a CS degree, I ship production apps. And I still, regularly, deliberately, code without AI on some of my work.
Yesterday’s client project was one of those. I could have had Claude Code read the whole repo and propose changes. I did not. I read it myself. I made the changes myself. It took longer. I did it anyway.
Here is why. The muscle that reads code, that holds the whole system in your head, that notices when something is slightly wrong before you can articulate why, that muscle atrophies if you don’t use it. I can feel it. When I spend two weeks letting Claude Code drive, I come back to a codebase the next week and it takes me twice as long to get oriented. When I deliberately spend days writing by hand, I stay sharp.
My CS degree forces this on me. Half of my assignments don’t allow AI. It’s the reason I can debug my own production code when Claude is hallucinating, or when my client is texting me about a feature they want to implement, and there is nobody to ask but myself.
If you are learning to code right now, you do not have a CS degree forcing you to build this muscle. You have to choose it. Nobody is going to make you. Every tool and platform and course is pushing you in the opposite direction because the opposite direction is where the money is.
The honest why
I care about this specifically because the people I see learning to code in 2026 are not Stanford CS students with a safety net. They’re career pivoters. They’re women in tech who already feel like impostors. They’re people who saved up to buy a laptop and are betting their next year on this working.
The AI-first learning industry is selling those people a shortcut that is not a shortcut. It is a detour that feels like progress and ends somewhere worse than where they started. In two years, when they try to get hired, they are going to find out that “I shipped ten projects with Cursor” does not mean what they thought it meant.
I don’t want that for you. I went through sales to CS to shipping client work, and I know how narrow the margin is. The thing that makes the difference is not speed. Its substance. The engineers who will matter in 2026 and 2027 are the ones who did the boring, unglamorous, AI-free work of learning to code with their own hands first.
Do that. Use the ladder. Turn the AI off for thirty days. Write bad code. Fix it yourself. Get mad at it. Come back the next day. In six months, you will be a different person, and in two years, you will be the junior everyone is still hiring when the rest of the cohort is stuck.
Let’s Build It Beautifully,
Fab