Comment by officialchicken
8 hours ago
It works fine for webapps and other slop-adjacent projects.
If you try to do anything outside of typical n-tiered apps (e.g. implement a well documented wire protocol with several reference implementations on a microcontroller) it all falls apart very very quickly.
If the protocol is even slightly complex then the docs/reqs won't fit in the context with the code. Bootstrapping / initial bring-up of a protocol should be really easy but Claude struggles immensely.
> (e.g. implement a well documented wire protocol with several reference implementations on a microcontroller)
I have had an AI assistant reverse engineer a complex TCP protocol (3-simultaneous connections each with a different purpose, all binary stuff) from a bunch of PCAPs and then build a working Python server to speak that protocol to a 20-year-old Windows XP client. Granted, it took two tries: Claude Opus 4.1 (this was late September) was almost up to the task, but kept making small mistakes in its implementation that were getting annoying. So I started fresh with Codex CLI and GPT-5.1-Codex had a working version in a couple hours. Model and tool quality can have a huge impact on this stuff.
I just vibe coded a VST. Runs a mix of realtime DSP and ML models. Really nontrivial stuff. It does exactly what I want.
Claude Opus 4.5 is truly impressive.
That's an app (running ML), not a protocol.
I hear people report the opposite.
The sloppier a web app is, the more CSS frameworks are fighting for control of every pixel, and simply deleting 500,000 files to clear out your node_modules brings Windows to its knees.
On the other hand, anything you can fit in a small AVR-8 isn't very big.
Whatever you do, your mileage may vary.
Yep, but I don’t intend to let that happen to my web app! It’s not that big and I intend to keep it that way.
Dependencies are minimal. There’s no CSS framework yet and it’s a little messy, but I plan to do an audit of HTML tag usage, CSS class usage, and JSX component usage. We (the coding agent and I) will consider whether Tailwind or some other framework would help or not. I’ll ask it to write a design doc.
I’m also using Deno which helps.
Greenfield personal projects can be fun. It’s tough to talk about programming in the abstract when projects vary so much.
I've been working with an agent to make a web-based biofeedback "application" which is really a toolbox of components you can slap together to support
and even do things like a 2 player experience. The code is beautiful, pure CSS the way it was supposed to be, visualizations with D3.js. I do "npm install" and can't get over the 0 vulnerability count. It's coding with React that's 100% fun with none of the complaints I usually have.
Given the amount of Arduino code that existed at the time LLM's were trained, I would have to agree that AVR-8 might be fine. For now it's on the Cortex-M struggle bus.