While the holidays were on I vibe-coded anything I could get my hands on.

While the holidays were on I vibe-coded anything I could get my hands on.
First, a couple of inputs:
- I last properly programmed something into production a long time ago, like 10 years ago probably — last time I wrote something by hand from scratch and all the way to the end as a ready-to-use project, seriously
- I know the syntax of a lot of languages, I understand all the basic programming principles (all that OOP, functions, even remember how monads work)
- but I don’t write code, though I can read any algorithmic trick, but I can’t do architecture stuff (it eats too much time)

What I tried:
- a frontend for a reader a-la
https://alphaxiv.org/
- make the exact same thing as a Mac app
- did various connections to our framework, like a plugin for Zed to use models that run through our uzu inference engine
- modifications of different repos to support local inference, for example here
https://github.com/alexzhang13/rlm
- solving various tasks in math analysis and olympiad programming from Kvant magazine (well, whoever remembers will get it)

In the end:
- frontend is a disaster for me, I straight up don’t know web programming at all, so I just can’t get what I want, the interface keeps sliding around, I fought for hours just to align stuff to a grid or get PDF previews so nothing shifts
- the Mac app turns out ok if you remove the interface; with UI it’s the same as web, but stuff like menu bar plugins that search articles by my queries — it does great
- any modifications and connnections, plugins somewhere, and simple API support, whether it’s our engine or something else, it does reliably almost from the first try, especially if you ask in advance to write an
AGENTS.MD
file in every repo so it’s easier to glue them together
- it solves problems head-on pretty well, but if I try to write a bot for some game or trading in a game, basic math is a mess. Mean is median, median is mean, and so on. Primitive mistakes, and fixing them only works by hand. If your chain has a bunch of computations, it just can’t carry it through to the end. It only works if you force it to write tests for every step. But imo for these tasks that’s overkill.

Bottom line. The progress is very noticeable. Codex, in my opinion, is smarter for one-shot stuff, but Claude code is for projects that are wrapped in tests and md instructions at every step. For now it feels like the stage of writing integrations, connections, basic things and review is close to automation. But architecture and math in code are hard. UI I’m not even gonna talk about — either I’m doing something wrong, or it’s completely busted and it can’t improve anything beyond what came out in the one shot.

I use every day:
- Claude code with plugins, type the command /plugin and there’ll be a sea of useful stuff
- cursor
- codex via the web client

Links from the original post:

• http://AGENTS.MD/