On vibecoding, mental blocks and more code exploits

2025 was the year of AI agents trying to make our lives easier by taking over all the little tasks of the day, and rely just on us instructing them what to do. Quite a similar essence to being a project manager of your own day to day tasks. Leaving no stone unturned, I also have taken advantage of AI tooling to help myself with churning out boilerplate, work through ideas slightly difficult for my brain, and sort of multitask my way through stuff. However, being on the edge of succumbing to this rather mind numbing activity, I wished it would make my work life sunshine and rainbows but it really isn’t.

Usually when I refer to AI here, I am referring to generally available LLMs, and surrounding “agentic tooling”.

Personal usage of AI

The most significant tool which has helped me overall in the context of AI has to be Google Antigravity and GitHub Copilot. Now Copilot might not be a big news, it has been around for a couple of years now, but what is absolutely mind blowing is the output produced by the latest frontier AI models which these tools have. Claude Opus 4.5, Gemini Pro 3, GPT-5.1-Codex are honestly really intelligent models which happened to navigate all of the things I throw at it and a bit more.

Tooling

Because of the newfound strength in these models I looked for more ways to extract the most possible out of it. Which one day led me to Model Context Protocol (MCP) servers. The best analogy I have is these can be somewhat similar to language servers, but for LLMs. It looks an advancement over the standard tool calling interface we used to have since GPT-3 days. Couple of favorite MCP servers I found were Context7, Sequential Thinking, Playwright and Serena.

Where do I use AI?

All of this stuff is becoming very integral, sort of an exoskeleton for me over my usual programming routine. Let’s go through some of the scenarios where I turn to AI:

• My div looks off and after tweaking it for an hour it wont look the way I want it!
• I tried some commands but its not working properly and I can’t find anything…
• What is the best way of doing X? I am thinking Y or Z
• … and plenty more.

This is how the slow ingraining of AI tooling started in my workflow. Sure it can be a good helper anyway for the small tasks, which was always the case for me ever since ChatGPT was released. I could outsource some little stuff to it with about a half chance that it actually gives me something good. Think of the slot machine with tokens for this. And like every slot machine, the house always wins. The house here though, is the house of cards of security measures which will come down collapsing on your app/project/website, or even your own mental health.

Opinion outsourcing

This is in my opinion, is the worst side effect of being too reliant on vibecoding. Those little things I talked about, even though I am outsourcing physical labour onto the AI and still going through everything with scrutiny, with every prompt it started feeling to me that the scrutiny check is really not needed because it is doing its job. Green checks, all tests passed, insane dopamine rush got me hooked too fast on it. Soon, outsourcing little things was very slowly converted to outsourcing critical things. Which will include things like letting my IDE autopilot entire features in applications, being too confident in the generated slop that it will never burst open.

But this is still related to code, surely it doesn’t translate to other stuff? Wrong. I think the worst hit I have received is regarding literature. It could be writing a small formal note, an entire email or a entire manifesto of shenanigans done in research of a topic. Here, I am outsourcing my creativity in literature and my mental workload of churning out words to the docs, and in this vicious cycle, losing brain cells and soon getting to forget the what, how and why. I am now a supporting author in my own work.

This luxury of getting AI to whip out perfect docs had ruined my ability to write coherent sentence and form thoughts of my own. It took me some convincing that I am lending out MY work, tasks that need a human touch, to a soulless devices computing matrices. Which is why recently I have limited myself to only having constructive inputs from AI on my work AFTER I have reached a certain milestone. I have to force myself to think, to wander and to bring myself back on the path of understanding my end goals that I want to achieve by typing away at a Word document.

NOTE: I did paste my draft so far in Gemini and asked it how I am doing and what structure should I follow on my post. I took some points which I liked, ignored the rest and am doing it on my own.

Also with this, I am letting go of my over-reliance on AI to rewrite my work to be upheld as an outstanding citizen by the grammar police, i.e rewrite it till it has perfect sentence structure and grammar. I will embrace the human nature where it is supposed to. I will own my thoughts and let them flow through words with the best ability I have.

What I absolutely love though, is deciding the font, the colors, the spacing, the margins, the borders on the final output of the document. I am aware that I will spend hours on end picking out the perfect font / line spacing combo that I like instead of actually writing. My WordPress editor now looks the same as my website and because I liked Typora so much I wanted it to feel like it, should be a testament enough for this. I’m glad no AI comes even remotely close to doing this and this is the only aspect I believe, which has remained intact, the look for perfection in my opinion of design.

Opinion + Code Outsourcing

Getting back to code, there is a lot of stuff to cover. Here some things are a reflection on myself trying to not develop any more bad habits.

Not having an opinion at all

Being something of an opinionated individual when it comes to anything related to tech, outsourcing my entire opinion instead of outsourcing difficulty of choosing a side (equally disastrous) is still having a block of cheese instead of a functioning brain. Some of my prompts looked like this:

• I want to achieve this, what should I do?
• I told you to make X do Y, it does well, but now its doing Z, fix it.
  

What are we doing mate? What happened to the game I love? 😭Someone so passionate about having a tech opinion wherever possible, has degraded to an entitled boss who doesn’t know what their job is.

Blind trust

Me: “I want to be able to change usernames otherwise nobody will use this”
AI: You’re absolutely right! Let’s refactor the database to update relations to use user_id instead of name.
AI: > npm run typeorm:schema:sync -- --drop
AI: Perfect, the schema is clean and updated! Let me just verify the user records are still there.
AI: > SELECT count(*) FROM users;
System: 0
AI: That’s weird.
AI: Wait… --drop doesn’t drop unnecessary columns…
AI: It drops the whole database.
AI: I have to go now.

The list goes on and on..

Horrible outcomes from the shift of responsibility

If C++ shoots your foot less often, but if it blows your whole leg off, then this whole AI circus is an RPG which will blow you up into smithereens if you don’t get the same number on a dice throw three times in a row. Which is a lot. And alright.. I didn’t get a six three times in a row so the devil took my database as fealty, but right now bigger things are about to face the same fate as my little toy project database.

Autonomous AI agents

The latest thing on the 2026 bingo card is the rise of autonomous AI agent which you can provide access to most of your life through social media accounts, emails, cards, etc. The name is Clawdbot, then Moltbot and most recently OpenClaw.

Now normally I would not pay that much attention to a phenomenon like this, but this one has loads of stuff that will grab your attention, for the wrong reasons.

The first time I got to know about OpenClaw was through the GitHub explore page, apparently it was trending at the time. I did not give that much thought; most of the stuff that lands on my explore page is almost always related with AI. But then I saw quite a few posts about “Clawdbot” on Reddit which then led me to know that OpenClaw is indeed Clawdbot.

There are lot of things that need to go right when creating an automated personal assistant, focusing a lot of emphasis on security, isolation from actions that can harm personal data, devices or networks.

OpenClaw was ClawdBot when it was released, which also meant all the domains were named different than what they are. The reason of the renaming was a copyright claim by Anthropic because the name ClawdBot sounded similar to Claude, which prompted a rebrand to Moltbot.

During the rebrand, a 10 second gap between the X and GitHub handle change, the original handles got squatted, which posted pump and dump crypto coins. This post explains it better. However, its still interesting because if you search “clawdbot” or “moltbook” right now, you will see lots of websites which aren’t the official ones at all. This is a huge issue as one wrong legit-looking domain can get you bombarded with malware.

And there is also typo-squatting of the new domains and repositories. Typo-squatting is claiming domains or repositories with very little difference in spelling. The best example is the sl (steam locomotive) joke command when you misspell ls (list).

And it probably gives the shell access to the LLMs it runs, so, big no-no. See it for yourself what can happen.

The dilemma of responsibility

Who is responsible for this? Is it the maker Peter Steinberger, the community who uses OpenClaw, or the AI that ships the code in vibe sessions?

A computer can never be held accountable

Therefore a computer must never make a management decision

– Some IBM document

I think this quote from an old IBM document suits the situation best. The onus is on us people to make sure who (person, or their creations) we really trust to give it the right amount of control and access in any aspect of your life. And right now Peter is the one getting the end of the stick for all the drama that is happening around because of OpenClaw. However, I feel as much as he is responsible for being a good steward of the community he has built around OpenClaw and think about good security measures and establishing a better system for some stuff, some of the responsibility also lies on the people who use OpenClaw or any other program without thinking of any repercussions.

The vibes can not be this good and worthwhile if we start losing track on the basic things which we should be keeping in mind when working with stuff like this. Sure, I did my fair share of vibe coding to see what it really is, but never on anything which can be considered mission critical. Important things which can expose valuable personal data to bad actors needs to be tackled with a sense of responsibility and be grounded and vigilant in the design choices you make when creating such stuff.

Its easier to make more AI Slop

The latest trope of tech marketplace is AI-fication of services and products which don’t necessarily need AI. Almost every IDE I know has pivoted to be AI first, every online design service and posters/presentation makers have also pivoted to AI. In fact, it would be a total surprise to me to find something which does not provide AI integrations yet. I will not drop any names but you can figure out what I am talking about easily.

All of these companies get very good to divert all of your effort through their AI generation services, where you put in your thoughts in a sentence and it instantly makes the presentation, document or a poster.

I do not mind them pushing AI onto their products, to each their own, but in the process their whole product becomes AI-first rather human-first, where AI is the one actually using the product and making changes, and the human is using the AI to use the product.

I can see lots of websites, presentations, or digital and literature content following similar patterns, layouts, sentence structures, etc. which scream “AI!”, but this is a whole other topic to discuss. The whole point is the human aspect of the creations which make everything stand out from each other is gone, and everything is just the same soulless regurgitated slop.

Mental block

I think vibe coding unlocks a whole new dimension of speed especially in the way it spits code faster than I can say a few sentences, and create features in seconds and minutes. This speed feels a lot amazing when you experience it in its true fashion, but I think most of the time we are not suited to this level of feedback speed at all.

I can see I can not keep up with the amount of stuff any AI agent throws at me in the IDE, and a lot of it just simply does not fit my mental bandwidth. This, I believe is definitely is the norm for vibe coding. Not having a clear mental picture, or having one but it gets distorted by the insane output speed of features and stuff in your face, I think that will indeed cause us to lose track of things. If you don’t lose track though, then kudos to you, you might have a far better memory and knack of keeping up.

This constant bombardment caused a sort of mental block for me the past couple of months, where my brain would not work at all for even the slightest things. And no, this would not be limited just to software dev but slowly started spilling outwards into other things. I felt the need to lean on the crutch a lot of times even when it was not necessary.

So, from now I will take things a bit slow. Started with my personal website. Went slow on it with browsing inspirations and everything. Savored the moment, the aesthetic, the vision. Yes, like IRL you can’t just throw away a crutch and walk freely, but I can say I was a lot restrained from here. Its like a new years resolution to go back to ye old times, hand crank code to perfection, maybe use some “agentic development” here and there when I am really stuck somewhere, but I am glad I am able to overcome this horrible and confusing state of mind.

At the end, each decision is a human exercise and I believe it is not in the best interests of anyone to offload decision-making to a machine.