JSNation is coming soon, and besides my talk (I’ll drop the link in the comments so I don’t spam you with it for the tenth time 😅), there are also ...
For further actions, you may consider blocking this person and/or reporting abuse
I wrote a short post about the same topic a few months back: How Much AI-Generated Code Are We Actually Shipping to Production? My Reality So Far and my experience is very similar.
For older codebases, AI often isn’t that useful because understanding the system matters more than generating syntax.
In newer codebases, though, AI works much better when the architecture is already well established and best practices are being followed. In those cases, it can generate similar code that fits the existing solution quite well. Usually, it gets me about 85% there, and I fix the remaining 15% myself.
And then there’s the third category: hobby projects. There I use AI for almost everything 😅
But we’ll see how token costs evolve over the next few months. Maybe even hobby projects will push me back toward good old manual programming 😂
This is a fantastic post, and honestly, everything you describe fits together into a very logical pattern.
And yes, I totally agree about hobby projects too. That’s where Anthropic servers start catching fire 🔥😂 In personal projects I use AI for almost everything.
Although… I already noticed my own behavior changing a bit because of token costs 😄 A few months ago I would happily generate even boring boilerplate through an agent. Now I sometimes catch myself thinking: “You know what, the old-school CLI command is probably cheaper and faster here.” So maybe I’m already adapting to the token economy and it won’t be that bad after all 😂
And btw, I’ll actually be in the Czech Republic in October for FrontKon, so you should slowly start preparing a list of good Czech beers for me 😄
Yep, I think we’re now entering the phase where people will start learning how to use tokens more efficiently and maybe the whole AI bubble will stabilize a bit because of that.
And nice! As the leading country in beer consumption, the list of good Czech beers could get very long 😄 Not sure the DEV comment section is big enough for that 🤣
Apparently we’re already reaching peak absurdity, because I recently heard about LLMs being used to improve prompts so that other LLMs consume fewer tokens 😂 The ecosystem is slowly evolving into AI agents optimizing other AI agents.
And yes, Czech beer absolutely rules! Unlike those mysterious dumplings (knedliki) you people keep trying to convince the world are food 😂 Although to be fair, the calories saved by avoiding knedliki can then be responsibly re-consumed in the form of vitamin B-rich Czech beer, so the nutritional balance eventually restores itself 😎😂
Agreed, it’s getting ridiculous.
Oo, so you’re already an expert 😄 Knedlíky are the best side dish! Especially with goulash or svíčková, and after that, 3–4 Czech beers… that’s perfection 👌😆
And calories? Well, in the Czech Republic, you’d rather not count them when it comes to food and beer 😅 But I think it’s similar in Poland 🤣
Go with some of the best free provider options via
Kilo Code
Cline
BlackBox AI
At work, I am mostly the person saying, "let's not use LLMs for this, it's a terrible idea because..."
I have two current code projects, and one I'm doing myself, the other I'm using Opencode for. The Opencode one is significantly worse and requires me to step in and rescue it a lot. I also have to keep asking it to explain what it thinks it's done, and to tidy up things that are left behind and to not keep reinventing the wheel. Oh, and I have to start every session with, "for the love of god do not use tailwind this time".
What I am using AI for is asking questions about code, and for admin tasks like creating a changelog from the git history of a project and linking it all up with Jira, that sort of thing. I mean, you still have to spend time making sure it's not coming out with something that's actually bonkers batshit but it does save a lot of time overall for those sort of chores.
This matches very closely with what I’ve been seeing too: context is the key thing here, and no agent or model is really a silver bullet.
My experience with coding agents is very similar. They work best for relatively simple, well-scoped tasks like: “add a small feature HERE and HERE”, “split this component into smaller ones”, “generate some boilerplate”, etc. And honestly… even there they sometimes manage to completely derail themselves 😅
And BTW, as you already know, I actually like throwing Tailwind into hobby projects 😄 But I usually have to explicitly ask for it, and specify the exact modern version too, otherwise the agent starts building some terrifying spaghetti CSS monster inside a random global.css file 😂
The legacy-vs-new split in the comments matches what I've found, and I think the dividing line is sharper than age. The new codebase where it gets you 85% works because the context is in the code: the folder names, the types, the seams all tell the agent what's going on. The legacy one collapses because the context is in people's heads and in that file a junior wrote seven years ago. What's changed how I work is treating that as a forcing function. The stuff I used to keep in my head the agent can't read, so it has to go somewhere it can: clearer names, tighter types, a short doc of the weird bits. It helps the next human too, which is the part that was always worth doing anyway.
Wow, thank you so much for this comment. I honestly suspect you’re exactly right.
To be fair, I’m currently not working with a “well-written legacy system” 😅 My project has a lot of that classic tribal knowledge situation: one senior got moved to another project, then a junior got assigned, somebody hacked something together, somebody left a TODO in 2019 and disappeared into the void 😂
So yes, I think you’re absolutely right that if the codebase itself communicated the context better, the agents would probably perform much more effectively.
Which immediately raises another interesting question for me though: what will happen to all the projects being created right now with heavy agent usage but without much architectural thinking behind them?
At some point, somebody will have to maintain all of that 😅
And honestly, I already saw someone comment somewhere that maintaining AI-generated legacy systems in a few years might become a literal gold mine for experienced developers 😂 What do you think about it?
If you're interested in my talk at JSNation, you can register here and watch it for free 😊
Here it is
The talk is called “Rewrite or Refactor? How to Safely Move Legacy Apps to Modern Frameworks” 🚀