How To Use AIs Incorrectly (Comprehensive Guide)

You are probably using them wrong, too.

In the age of Artificial Intelligence, everyone started to use them. I’ve been an OpenAI user since the GPT2 times (waaaay before the OG ChatGPT), and I’ve seen them come up with clearer, smarter answers over time. Still, if you know how they are functioning, you will stop using them incorrectly as well.

how they function.

There is a mathematical formula behind every LLM out there. It doesn’t matter how complex they are, how many layers there are, how much data it is trained with; there is always a formula. Computers don’t “think”, they calculate. If you expect to get everything out of it, it will fail. Try running “Build GTA VI, make no mistakes” the next time you use Claude.

I’m not going to take you into the intricate details of how they function under the hood, but let’s oversimplify them with the following formula:

Where Q is Queries, K is Keys, V is Values, QK^T is the dot product between all queries and keys, the square root of d_k is the scaling factor, where d_k is the dimension of keys, softmax is a function normalising the score between 0 and 1, and multiplying with V blends the actual semantic values (meanings) together based on calculated attention weights.

This calculates the context and tries to get the next token after the current one, and so on. Obviously, there are more “formulas” than this alone (like token ordering, training formulas and such), and if you are old enough to balance a stick with AI yourself, you would know a thing or two about these formulas.

training data.

AI, is stupid. Let’s agree on this one. I can speak for myself: If I’ve created something that is too powerful, I won’t publish it on the internet. Some things are just pure “experience”. AIs, however, are trained on StackOverflow, Reddit and any information that the developers could find. And the information you can find online is “average”. I can summarise the knowledge of AI in general as follows:

It is a wonderful tool to get the very average data, very very fast.

It is a glorified search engine, that is what it is. You cannot build something that is not on the internet, published, or trained by the AI developers before. It can guess, but without correct information, it will guess pretty badly. You can do it by hand quite all right with enough time, but using AIs makes the knowledge-acquiring process faster.

what i mean by “incorrect usage”.

I’m not against using them for work; what I mean by that is purely depending on them all the time. Most of the time, you are under-feeding the LLM by giving it the least information possible, getting a result in 30 seconds, and spending 2 hours to debug and make things work. If you can’t find information online with the parameters that you give AI, you won’t succeed. I’m using Antigravity and Gemini myself most of the time, but for repetitive tasks. I’m not depending on it; I’m using an eclectic approach. If it goes dark one day, I’ll continue creating more projects without relying solely on it. I micromanage it all the time, anyway.

Imagine describing something you want to be done to an experienced programmer. If you give the same prompts to a human, they surely will come up with more questions than answers. How, where, and why questions are human, plus we can say “That’s not going to work”, if you are experienced enough. In the case of LLMs, they mostly accept the first interpretation they understand and proceed from that. Even though you see what is going to happen throughout the planning process (in the case of Antigravity), and even if the plan looks spot on, LLMs lie. It changes stuff that it shouldn’t be messing with. If you don’t check what changed and don’t understand your code better than the AI you are desperately depending on, you won’t even know why the issues come back to the surface. After all, it was working fine just 2 prompts ago. Then, a 2-day debugging process begins.

it’s repetitive, not creative.

You’ve probably seen those “ok, let’s make the website with AI” websites; they all look the same. When I create something, I’d like to add my salt and pepper to it. I make it mine to begin with. Remember, you are not only making something; you are making art. And you are the creative art director, the coordinator, the make-up artist and the costume designer, and everything in between. When you let AI make those creative decisions, it will fall back to its “average stupidity” and give you something very average-looking, which was agreeable upon on the internet in 2024. It suffers from “Fallacy Tendency”. Unless, again, you give it a specific direction and intricate orders to follow. Still, it won’t create the new GTA VI for you, but it can give you the camera controller for it. Be specific and don’t expect much, especially for the things that you have no idea of.

“But Yağız, everyone on the internet is vibe coding all the time and creating companies based on those programs. How do they do that then?”

Well, let me introduce you to the next topic.

it is not what everyone thinks it is, even their own creators.

Big companies laid off their developers en masse, and they realised their own mistake. They thought their tools could replace seasoned developers, but here is actually what happened (Joberty, 2026):

• AI-generated code produced up to x1.7 more bugs,
• Maintenance workload increased by around 38%,
• 96% of developers don’t trust AI-generated code (including me).

They are silently rehiring their seasoned developers to fix the mess that AIs have created over time. Because of this fact, which I also agree on: Faster code generation doesn’t mean better code.

It is not a competition between humans and computers; it is a tool for us to use. It is a great tool to make automation processes faster, like debugging, unnecessary typing, repetitive tasks and so on. It is not a generative tool as everyone thinks. They initially came up as a tool, but with time, everyone thought that they could replace humans. Delusional, if you consider the performance of it, today (I expect it to be much more capable in 5 years, by the way). If you are above the average, like the information that AIs are fed, you are safe.

then, how to use them properly.

As I described before, if you use them as they are designed, AIs will be very beneficial for your workflow. I highly suggest that you to create a design template for your AI to make your new projects easier to work with. That is one of my example templates that I often use for my projects (you are free to use it yourself):

# Next.js-Based Projects Rulebook
1. If we are using pnpm in the first place, use pnpm whenever you can.
2. Unless asked to do so personally, skip automated visual checks from Chrome, since that slows down the process.
3. If we decide to use Framer Motion in any part of the coding, don't use the transition-all property. Instead, use something like transition-[opacity] for the things that Framer Motion don't handle. That breaks so many things for the long run. Also, try to use Framer Motion for the rest of the animations for that element, if we decide to go with Framer Motion, of course.
4. Blur changes and blur in general break the Safari experience for the end-user, and also slow down the website. If we are trying to animate or use blur in general, notify the user about the situation and suggest alternatives instead.
5. Don't run the pnpm build command to verify the integrity, as it sometimes breaks the pnpm dev command that is active. Use other commands for linting and TypeScript checking.
6. If we have an animating object with any means, we must write the "will-change" property for that element.

Those are things that happened to me more than once for my projects, and I’ve added these parameters to prevent them from happening. All things considered, Antigravity follows it, mostly.

The other thing is the model that you are using and how you use it. I never give a “chat” more than one task, and tasks are generally pretty light. I don’t even use the Pro model; most of the time, I’m using Flash (Low or Medium, rarely High). Some tasks require Pro to create a good plan, and even some of the hefty tasks that I create can’t be achieved with that, either. In that case, rather do them by hand, or divide the task into small steps. On the topic of plans, always read them, comprehensively. Create comments on topics that need changing, and affirm the areas that need to be confirmed. You are holding its hand, not the other way around.

When your plan is created and sounds nice enough to fulfil your needs, go ahead and approve it. While the AI is working, monitor it. If you see something going catastrophically wrong, stop it. After the code is generated, always double-check the changes and understand what is going on. Use your experience and critical-thinking skills to identify any future problems. If the search function can fail under immense tension (prompting the database every time a keypress is identified), expect it to fail on a Black Friday sale. You are the one responsible for that implementation; you will be punished when everything goes south.

Don’t make the AI fix small things like layout changes, text centring, and such. You have the capability to fix them yourself. It is not a repetitive task; it is just a task. This will both reduce your bills at the end of the month and give you more control. Always have situational awareness and never push something to production that you don’t understand and trust. Imagine this scenario: Would you trust someone else’s work without checking it twice to push it into production blindly? How much trust do you have in someone else’s work when your own job or product is on the line?

If you follow this structure, you will get helpful results, more or less. Don’t push your luck with the things AI is clearly not capable of. Average stupidity can be helpful in some situations, especially if you think of the AI as a “very inexperienced junior developer who has a fast internet connection”. With enough time, everything is possible, and possible because of you, not the other way around.

conclusion.

The AI is developed to be anyone. If you don’t want to be that “anyone”, you should have a plan, a creative direction and a vision that overcomes the challenges that come with using those oversimplified tools. No matter how powerful they become in the future, they simply don’t understand what a creation feels like. Even if they do, my understanding of beauty and yours are two different things.

For example, I created my portfolio (iskirik.com) with my beauty concepts in mind, but it can be too artistic for your taste. The WebGL rainbow at the footer of the page alone took 9 sleepless hours to finish; the AI basically wasn’t capable of doing that, so I created it myself. The background is a 4D Voronoi texture that both evolves with time and the 4th dimension changes with the scroll location. Again, the AI couldn’t do that, so I took over. It is not “vibe-coded”, it is structurally designed.

If I can leave you with one piece of advice to carry into your next coding session, it’s this: Never let an AI write code you couldn’t write yourself, given enough time.

Treat the AI as a brilliant, lightning-fast, but incredibly naive assistant. Use it to clear the brush, to handle the boilerplate, and to fetch the standard templates. But when it comes to the architecture, the critical edge cases, and the soul of your project, take the wheel.

In a world increasingly saturated with average, AI-generated noise, your value isn’t measured by how fast you can hit Tab to autocomplete. Your value is your taste, your critical thinking, and your willingness to spend 9 hours on a WebGL rainbow just because it makes you feel something.

Don’t outsource your thinking. Be the engineer, not the prompter.