Technology makes us lazy.
That's not an opinion but a fact. We can't do mental math, find addresses, or memorize phone numbers anymore. That's the problem with relying too much on a piece of tech. Smartphones, I'm looking at you.
The same thing happens in coding, with AI and vibe-coding.
I'm guilty too. I've been experimenting with AI to offload my plate of boring tasks. And when I can't think of an answer immediately, I'm tempted to go straight to the genie in the bottle to grant me a coding wish.
And I'm not alone in this. The other day, I found this question on Reddit,
"It's been a while since I coded on C#/Unity so I'm very rusty on the concepts and relying too much on ChatGPT makes me feel like I haven't learned anything and can't write code on my own without doing basic mistakes... How do I learn everything back? Isn't there a way to refresh my mind? A really good video on YouTube or something? I want to stop using AI and code on my own."
For the original poster and anyone else who wants to break free from AI, here are 10 ideas to try:
0. Ban AI.
Think of AI as calculators in math classes. You can't use them until you know the procedure you want to automate by hand.
Like any mom disciplining her kid, "No more AI until you do your homework..."
1. Study your main language syntax.
Get to know the syntax of your language of choice: write variables, functions, loops, classes...
For that, grab a textbook or watch a "all you need to know about X in 4 hours" YouTube videos. But don't just passively consume them, recreate the examples and projects from them. By typing them out. No Control C + Control V.
2. Know your standard library.
Get familiar with your standard library:
- Write a variable and see what your editor or IDE suggests.
- What methods can you use with that type?
- Look at their signature and docstring.
3. Study SQL.
No matter how powerful ORMs are, we can't escape from SQL.
We've had SQL since the early 70s and chances are we're still using SQL another decade or two.
Learn to create tables, write queries aggregating results, and learn about JOINs. Download a light version of the StackOverflow database and play with it, if you want realistic examples.
4. Build a toy project from scratch.
OK, I'm not talking about reinventing the wheel to write your own text editor or something.
I'm talking about building a recipe catalog, a todo app, or a CLI wrapper around a free API. And build it from scratch: right click, then create new folder in your editor or IDE, and so on. It will teach you a lot.
5. Find your own answers.
When you get an error message (you will if you follow #4), resist the urge of going back to AI or simply asking a friend.
Try to figure out errors and exceptions on your own. Start by googling the error message. There's a thing called Google that finds web pages with answers to our questions. Sure, it's old-school but it builds real muscle. Remember, AI is still banned. (See #0)
6. Learn the most common data structures.
80% of the time, you'll only need lists and dictionaries. But there are more data structures of course.
Learn to use them and how to implement them. You won't have to implement them from scratch at your daily job, but it will stretch your problem-solving muscles.
7. Study a textbook on Math for Computer Science.
Unless you're working on niche domains, you won't need advanced Math.
But grab a book on Discrete Math (or Math for Computer Science) and study a chapter or two. Again, to sharpen your thinking.
8. Practice rubber-duck debugging.
You will get stuck a lot. That's a feature of being a coder, not a bug.
When that happens,
- Grab pen and paper
- Go through your program line by line
- Talk out loud
9. Read the official documentation.
Pull out the Mozilla's Web Docs, Microsoft Learn, and any other official source for your language of choice, and not only read it, but come up with your own examples and think of how you can use what you're reading in your own code.
AI is a blessing for learning. Ask any veteran who learned from reference manuals, language specifications, and magazines, they'll tell you. Just don't let AI think for you.
OK, let's slightly lift the AI ban, don't use it to generate code. Use it as your copilot, not as your captain.
To help you build hype-proof skills, I wrote Street-Smart Coding. The roadmap I wish I had on my journey from junior/mid-level to senior.
Top comments (24)
Point #4 nails it. But here's the catch: toy projects expose the gap between "I can prompt" and "I can debug."
The real test? When your from-scratch app throws
NullReferenceExceptionat 2 AM and AI suggests 5 wrong fixes. That's where strong typing (C#) vs weak typing decides if you sleep or not.AI writes fast. Experience knows why it breaks.
Yes, debugging is a completely different skill to master...one more people are offloading to AI
AI is definitely super helpful for developers, it’s like having a caffeinated intern who never sleeps. But let’s be honest, its biggest weakness right now is confidence. It will happily answer a question it doesn’t actually know… and do it with the confidence of someone who wrote the textbook.
And then, if you ask the same question later, you might get a completely different answer like it had a philosophical awakening overnight. Of course, I get that there are technical reasons for this, model updates, access layers, all that good stuff. Still, it can be a little wild when the “expert” suddenly changes their story.
Sometimes it feels like AI would rather improvise than admit, “You know what? I’m not sure.” And honestly, we’ve all worked with someone like that before 😄
Oh, the classical "you're absolutely right" when you confront AI
The other way I had a season with an AI and I explicitly ban it to say that. I wrote:
Stop prepending "you're right" to your answers.It answered:Understood.And it complied from then on! XDI like these advice, but I'd say that you shouldn't be in that situation in the first place. You should not rely on AI that much that you feel insecure with the programming languages. AI is great for asking questions about code, or about the API, generate code for a possible solution to evaluate a design path... It's surprising that a programmer can feel insecure without AI, taking into account the relative small amount of time passed since their appearance!
Yes, Baltasar. It's like sending someone else to the gym and then complaining when we don't see our muscles growing.
Point #8 about rubber duck debugging doesn't get enough credit. I've been doing a variation of this where I explain the problem in writing before even opening my editor - like a mini design doc. Forces you to think through the actual logic instead of just throwing code at the wall.
One thing I'd add to this list: read other people's code. Not just tutorials, but actual open source projects. Pick a library you use daily and spend 30 minutes reading through its source. You start picking up patterns and idioms that no AI prompt will teach you. It's how I finally understood how async/await actually works under the hood in JS - not from a blog post, but from reading through a well-written library.
The calculator analogy is perfect btw. You still need to know what operation to perform even if you're not doing the arithmetic yourself.
That's one of the best way to improve our coding skills. I wrote a post about it the other day.
Thanks! A good idea among many :)
being in high school and learning how to code in this ai era is tough. most of my classmates are using ai to do websites and i am here tryingto learn. very tough and thank you for this read
Building websites is everything from letting GitHub pages render your README to developing and maintaining a complex corporate website with customer login, B2B shop and database logic. It's totally legitimate to use no-code solutions to make websites. An early success might empower people or cause Dunning-Kruger delusion.
You can do this with AI (or StackOverflow) as well. Start typing, describe and refine your problem. Mind that you spend expensive tokens (AI) or risk downvoting and deleting instead of an answer, so don't post too early. Review and refine your question and try to narrow it down to a minimal reproducible example!
The most underrated "hack" in 2026: Reading the manual also saves futile and costly search engine requests.
When building projects, use descriptive variable names, meaningful comments and commit messages! Write a README and try to understand, review and maintain it. That's your own private "official documentation" of proven code examples.
Use AI where it helps, especially to get things done quickly for customers.
That's the point. AI is not intelligent. But you can be.
Further Reading
Thanks Cesar @canro91 for the insightful and actionable post. I appreciate and support your advice!
The dependency problem cuts deeper than coding skills once you're running AI in production.
There are two kinds of AI dependency worth separating:
This post nails the first one. The advice is right: ban AI long enough to rebuild the mental model yourself.
The second catches teams off guard. Session limits, model deprecations, provider outages, sudden pricing changes — any of those can take down workflows you didn't realize were fragile. The teams that recover fastest are the ones who actually understand what their AI is doing, not just that it's running.
Same underlying principle though. You can't diagnose a broken AI workflow if you never understood how it worked to begin with. The muscle memory of reading your own system — tracing what's happening step by step without a black box in the way — is what makes the difference.
The calculator analogy holds either way: you still need to know what calculation you're asking for.
Great framing. I'd add one more recovery scenario: what happens when your AI team goes down, not just your AI coding assistant.
I run 7 AI agents as my actual business operations team — they handle finance, marketing, sales, engineering, research. When I hit my Claude Max session limit and the fallback LLM wasn't calibrated, I didn't just lose a coding helper. I lost my entire operations team for 40 hours.
The recovery wasn't "go back to writing code manually." It was: which business-critical tasks need a human fallback? Which can wait? How do I triage when 50 cron jobs suddenly have no brain behind them?
The dependency isn't just on AI for coding anymore. It's AI for running the business. The recovery guide for that is very different.
Great point! AI, tokens, and credit cards are the new bottleneck
Hot take: There's going to come a point where coding by hand just isn't how things are done in the real world. Just like it's often quicker to use a calculator -- if you can do it in your head that's great, but no one is going to give you a medal (or a raise!) for doing so.
The article is great and I appreciate the tips! I'm just saying that AI will eventually be responsible for the writing the vast majority of code in professional settings, and "dependence" is going to be less a "disability", and more a simple side effect of practical reality.
Good point!
I think the framing of "AI dependency" as a problem is slightly off. We're not "dependent" on calculators — we just use them for math. The concern should be: are you using AI as a thinking substitute or a thinking accelerator?
If you can't articulate what you want the code to do before asking AI to write it, that's a problem. If you know exactly what you want and use AI to skip the typing, that's just efficiency.
The key diagnostic: can you read the AI's output and know whether it's correct? If yes, you're fine. If no, slow down.
You're right! We still need our judgment to tell whether AI is wrong
Some comments may only be visible to logged-in visitors. Sign in to view all comments.