Personal AI usage disclosure

This post is completely inspired by that one — in fact I did copy the starting paragraph verbatim as it spoke to me ¯\_(ツ)_/¯ (had to google “em dash” like a caveman to put it right here).

I agree with everything outlined there, 100% - so please read that, maybe after finishing this, because my personal history with LLMs is of course different.

My history of using LLMs

A better search engine

In late spring of 2026, my LLM usage boiled down to using Claude (which I have been paying for since 2025) as a ChatGPT replacement (which I have also paid for before, but I cancelled and switched to Claude at some point because I’d heard “it’s better for coding”) - just talking to it, asking for general programming knowledge and documentation, how to do this and that, basically a way better search engine, because SEO slop has ruined just googling stuff.

At most I’ve manually copied half a screen of code it made for me, like a single function, that I then rigorously reread and maybe also completely rewrote, and left a bunch of comments “yup this is megaslop, actually vibecoded this one, yes yes, I stooped that low”.

Actually trying out a harness

Then my employer kindly asked me, “how much of those infinite Claude tokens we recently started paying big boy money for are you using?” It’s nothing too bad, I do believe they were just looking into how much people are using/not using it to measure impact, optimize costs and so on. They are not forcing anyone to burn tokens or anything, my current employer seems more than fine in that regard compared to some of the stories we’ve been hearing.

At the time, however, I kinda did believe I needed to burn some tokens to not stand out from the crowd, so I actually tried using Claude Code then. I had heard of harnesses before: basically you wire a model to do certain actions via specially formatted tool calls. Instead of asking a slow and stupid human to read relevant files, research, debug and experiment, it can just, like, do it itself - you just have to run it through a program that allows it to. But I was kind of paranoid and skeptical about hallucinations and whatnot, so initially I did not really believe in the concept, despite Linus Torvalds himself saying he vibecoded some side-project by “removing the human obstacle between the model and the code”, which in hindsight makes perfect sense actually.

So I forwarded some simple tasks to Claude Code and then neurotically reviewed all the code it made. There was also this thing where I did not believe for some reason that models are good at copying things - like if you have some very specific string that needs to be refactored to another place, I did not believe it would reliably correctly copy it byte-for-byte. So I was meticulously checking every instance of that, mostly with Copilot, which I have been using for years.

The corruption

The thing is - everything it did (mind you, the tasks were simple) was flawless, refactor - done, I don’t like something, tell it to improve⁠/⁠fix⁠/⁠change it - done, implement a simple endpoint - done. And I see the reasoning, the steps it took; it asked for permission to run those little one-shot scripts LLMs do to test, experiment, reliably refactor something, etc. In the case of the simple endpoint, it looked around, saw other similar endpoints, and replicated the shape, the logging, code style, naming, you name it. And then the tests. Boy do I never write tests, at most I can maintain/extend an existing test suite - and it wrote a bunch of them. Yeah, some of them kinda test nothing useful, but it’s way better than nothing!

And so, as I was actually using the slop machine, not just seeing the absolute and utter garbage that image generation AI is used for - and, believe it or not, its output is not incomprehensible to me, you know, years of programming experience, decades at this point - I changed my mind.

To be clear, it is painfully obviously not perfect, nothing is, hello, but it’s good enough - often it is better than whatever I would’ve done.

Personal projects

Then there comes the productivity, boy, the productivity. At this point, I’ve already started using pi.dev for my personal projects, and that particular harness is peculiar in that the first vibeco.. cough-cough AI-assisted project you do with it is improving it. It is very minimal, but quite hackable with TypeScript extensions, and the whole philosophy is that you ask it to make those extensions to itself to better suit your particular workflow.

We made a statusbar plugin that runs a little curl command I copied from devtools to show me the usage limits. A tool that enforces a specific jj commit format. A sandbox tool that fires up a podman container for those one-off data processing scripts it makes, to not bother me with permission prompts. Oh btw, you want to install a permission extension; by default pi comes with none, meh.

Also after having a session you just ask it to add a point about something to the global AGENTS.md file so it always remembers some preference of yours, something you figured out you need to always make it do, etc etc.

I’ll give you a couple of examples: “Always use a tool/code to perform any mathematical calculation, no matter how trivial. You can use the calc command for this.”, “Always prefer to use uv when running Python or writing Python scripts” etc.

NixOS

And so after you get a workflow down, after you calm the paranoia down a bit, you can.. just go at it. I have not made any major NixOS config changes in months without the majority of it being Claude-generated.. and then reviewed, prompted to fix/improve/rewrite/change stuff, then squashed it all down into a neat jj commit. And the AGENTS.md of the nix config prompts it to try to build the config after it’s done, so if it sees issues it fixes those - at the end you always get something that at least builds. Which is why writing Rust or TypeScript is quite popular with LLMs, more automated guarantees earlier, kinda like that.

The thing about tinkering with NixOS, or just working on personal hobby programming projects in general, is that it takes time and energy to figure out stuff, which I kind of lack as of late. With an LLM, you ask it “yo can this be done” and after a bit of deliberation it will give you an answer that is significantly better than one you’d have gotten yourself after a day of research. And it didn’t just hallucinate it out of thin air, it can google stuff itself. And most importantly, what they do now is test. It will test its own claims before presenting them. It will look at actual sources from the nix store, be it nix modules or even literally the source code of the service we’re looking at. It will run a temp setup real quick to check if the suggested setup has any merit - and only then report its findings.

And then with all that juicy context already loaded, you just tell it to implement it, give it the choices you make and let it go at it.

Ghidra MCP

After installing Ghidra MCP and pointing Claude at it, I built a thing I utterly would never have managed otherwise: a mod that improves the Noita replay editor, so you can scrub the replay frame-by-frame, as well as export replays as mp4, which is a way better format than gif. It works much faster too, since we brought the real library in. Maybe I’ll write a separate post about all that.

Some context: some of the people reading this might know I dabble in Noita modding a little, in fact I made a whole framework that wraps around TypeScriptToLua to be able to write Noita mods in TypeScript. It has a built-in workaround for a certain dofile idiosyncrasy of how the Noita Lua API works, and a bunch of nice autocompletions for things, which was the whole point of making it in the first place, as well as many other utilities. I mostly made it for myself, it also has a CLI that can start an isolated Noita instance conveniently, build and publish the mod and so on.

And you can open Noita in Ghidra to look at the actual engine code - I’d argue Noita is one of the easier things to disassemble and look at, good learning experience. I have done so, annotated some stuff, found interesting bits and bobs.

Like, it would be physically possible for me to make that project without assistance, but the time and effort requirements would’ve been way past my limits for what is basically a hobby, especially with my low energy and mood reserves lately.

Review & Research

One thing that is undeniably harmless is asking it to review anything you wrote - yourself, no LLMs, big hater, sure. However, when doing some experiments with DLL injection for Noita modding, I just asked it to check the things I was doing, and it found countless bugs, holes, leaks and other actually embarrassing things I was doing there. Like, it would point you at things it thinks are important. And even if you’re a huge skeptic - even one out of ten findings that you, the person supposedly knowing what they’re writing, completely missed, sounds valuable to me.

That one I kinda agreed with even before this whole eye-opening experience I’ve described above.

Why it works

Unlike NFTs, large language models are unambiguously useful in a plethora of applications. They have been a major boost to my productivity and to my quality.

“But that means your quality was trash before them” - yes, I guess it means that, woe is me. For my purposes, an increase is an increase ¯\_(ツ)_/¯

And I have something that people who use LLMs to code without knowing what they’re doing do not - I understand what they are writing, I have the critical mass of knowledge (and opinions) where I learn from them and correct them, no actual vibecoding here.

Slop

All of that being said, there are two big things that are completely ruining the perception of AI, ignoring all the IRL material costs (RAM prices, datacenters, water for some reason, etc):

1. Generative AI - while it probably is still more useful than NFTs, 99 point idk how many nines percent of generative AI is literal cancer. Yes yes artists are crying, but also it’s just slop slop slop slop, tiktok slop, meme slop, 3D printing model slop, deepfake slop (ouch) and so much more.

2. Slop spam - people just.. misusing this amazing tool, to vibe-write reports, vibe-spam, vibecode shit and push the burden of review on maintainers. If I make a pull request with an LLM, I have a reason to, I am certain it does what it’s supposed to, etc etc. If the maintainer sees any clue of it being “AI-assisted”, they immediately close it, ban me and hate me forever because of hundreds and thousands of spam PRs by slop kiddies ruining everything for everyone.

The genie

You’ve heard this argument before, but - the genie is out of the bottle, yes, yes, they stole the internet, it is used for many stupid, malicious and slop purposes.

Refusing to put it to good use for your benefit does nothing.

And it does not mean you’re against somehow managing to get some justice for the aforementioned stealing of the entire internet - would be nice if they got some comeuppance, although that seems unlikely.

Conclusion

You see a lot of people on the internet fully embracing the slop - usually for social or financial gain. Pretending to be something you’re not also seems to be a thing, again literally either for social (ego boost?.) or financial gain (literal scam).

Also you see a lot of people rejecting AI outright, and while it makes sense for generative AI, for things like coding and some adjacent applications it is an incredible tool, invaluable even. You know, large language models seem to be good at this natural language thing, as well as coding languages.

And a lot of those opinions come from people who I am certain are exactly where I was before spring 2026 - they didn’t truly use a good LLM with a harness to work on personal projects, so they do not grasp how insanely good they’ve actually gotten, and how you can, in fact, herd them. And they denounce any project that used AI as “complete and utter slop garbage”, equating it to the slop-slop-slop-slop that (seemingly) most people are spamming.

Those are the two extremes, and both extremes are stupid.

Having experienced it properly, I am now in more of the AI-leaning moderate group I feel like.

It is sad; I hope that, as time progresses, more people will see that LLMs for coding actually have merit.

This did help readability, and the whole rant is still entirely my own.