Alright, I made a skill and gave my agent a complicated refactor to implement. They implemented permissioning, it was so huge that it took me days to review the changes. And having to go through line by line and optimize its slop code creates a lot of semantic satiation, where words lose all meaning
if you had to learn how to program you’re a scrub compared to me. i was born a coder. i just had to wait for my eyes and muscles to develop so i could work the keyboard. claude aint shit to me
The worst part about recovering from an accident is I’m going to lose my sweet sweet gains. Will I ever be able to dead lift two plates again?
A tail as old as time. Man versus car car wins.
You can tell that I use dictation for this because it’s spelled tail wrong
Ok, after a week of having agents write the majority of my code, I do have to admit, I am technically working faster, however there are a lot of UX, accessibility, and coding patterns that I need to constantly watch for. Which I could have caught the first time if I'd written the code myself.
So they say, as you use AI more you become more comfortable with letting the AI do more. So far my experience is that I'll let it do more, because I'm tired of doing code review and my mind starts to turn off. I don't think this is as effective as the model owners say.
Me: Oh I forgot to add `resolves #[issue]` to my commit that is now second from the head. I'm sure AI can add that for me.
AI: Proceeds to get stuck in git rebase hell.
Me: Yep...just like a normal developer...
i built an entire x86 CPU emulator in CSS (no javascript)
you can write programs in C, compile them to x86 machine code with GCC, and run them inside CSS
lyra.horse/x86css/
That is currently the solution. Temporal would make it a lot easier!
Alright, next week, I'm going to try and go all in on agentic coding. I'm still highly skeptical and I'm not going to be going all in on the vibes as I'll be expecting to give instructions to the LLM to refactor and not create pure slop. But let's see how much more *productive* I can be.
I'm working on a employee scheduling app called ScheduleLord. Having to deal with time and timezones is a special kind of hell. Why isn't Temporal done yet?
What I hate about AI is that it is always behind. I'm working on zod v4 and they keep recommending v3 methods. It's so god damn frustrating that they can't just go and look at the docs and find the information rather than rely on their outdated training.
When I say I don't know how to do something. I'm not saying, all hope is lost, let's give up (though, I might say that later). What I'm saying is that I don't have a frame of reference to make an educated guess about the problem, but I'm sure I can figure it out. And oftentimes I am able to.
I'm trying to think of a better way to write tests. A kind of next evolution of BDD. www.test2doc.com/blog/2026/02...
"...may the hashes be unique and the tests always green."
Sometimes LLMs toss funny weird shit in to my conversations when we're done talking. This is probably the funniest. I think it does this because I'm always making jokes in my prompts.
I just vibe coded up a feature for test2doc. I didn't just blindly accept it. I forced the agent to TDD and then we refactored the implementation to make the code better. It was quick, but I'm not sure if I enjoy coding like this...
I'm experimenting with vibe coding a bit on OMA3 and the LLM has gotten stuck in vim. Classic rookie mistake AI.
Last part of the RWSDK tutorial. www.test2doc.com/docs/tutoria...
All done!
Amazing blog post on ASCII art rendering.
alexharri.com/blog/ascii-r...
Here’s part 1 of this week’s 4-part Better Offline Enshittifinancial Crisis Special. I walk you through stage 4 of enshittification, when companies decide to enshittify shareholders, and the concept of stock itself.
podcasts.apple.com/us/podcast/b...
Linktr.ee/betteroffline
Got a new tutorial around RedwoodSDK and TDD, and also going a bit into SOLID principles and refactoring. www.test2doc.com/docs/tutoria...
It's funny that I literally thought this was going to be one of the more simple tutorial sections and ended up being maybe the most complicated one.
I agree with Claude & Gemini, but I was leaning toward GPT's solution originally. I find it weird discussing coding tradeoffs and best practices with these LLMs, because I find it hard to trust them.
So while working on my tutorial I got bitten by not following the Rule of Three (have three similar examples before abstracting code). And I talked with ChatGPT, Claude, and Gemini with how I should approach this. Claude and Gemini advocate for de-abstraction, GPT thinks about a middle ground.
I feel this is an apt metaphor on why it’s important to engage and change these people’s minds rather than stop doing what you know is right. Because we don’t want them to shut you down and take power. I’m very aware, easier said than done.
It's hard to understand what each foundational model is good at. Like I trust Claud with coding stuff, usually... Usually I find Gemini good at non-coding stuff that's still related to tech. But I was just asking about accessibility and ChatGPT was better than the other two.
Terrible app idea: a dating app but it only shows your proxy social proof, like car or shirtless bathroom selfie.
Released new part of my TDD tutorial. www.test2doc.com/docs/tutoria...
Also made a blog post summarizing my 2025.
dethstrobe.com/blog/2025123...
I think to prove that test2doc (my reporter the generates docs from tests) works, I think I need to prove it by doing something that requires SOC1 and 2 compliance. I think I'm going to try and make an online bank. I think I'm going to do it in public and open source the whole platform. Stay tuned.
I think I hate writing tutorials. But the problem is, I think it's extremely valuable to help me even understand my own best practices to explain it to myself.
And the marketing and theme of the site is 1980's synthwave meets dracula. Boom! Marketing writes itself. *drop mic*
