That it authored in the first place?
Actually, no. When LLMs produce good, working code, it also tends to be efficient (in terms of lines, etc).
May vary with language and domain, though.
Let the LLM do the boring stuff, and focus on writing the fun stuff.
Also, setting up logging in Python is never fun.
If it's a new, non-trivial algorithm, I enjoy writing it.
Honestly I don’t think customers care.
I have also a code reviewer agent in CC that writes all my unit and integration tests, which feeds into my CI/CD pipeline. I use the "/security" command that Claude recently released to review my code for security vulnerabilities while also leveraging a red team agent that tests my codebase for vulnerabilities to patch.
I'm starting to integrate Claude into Linear so I can assign Linear tickets to Claude to start working on while I tackle core stuff. Hope that helps!
Would you mind linking to your startup? I’m genuinely curious to see it.
(I won’t reply back with opinions about it. I just want to know what people are actually building with these tools!)
Oh, and the chatbot is cheap. I pay for API usage. On average I'm paying less than $5 per month.
> and I don't have to worry about random hallucinations.
For boilerplate code, I don't think I've ever had to fix anything. It's always worked the first time. If it didn't, my prompt was at fault.
I'm deliberately trying not to do too much manual coding right now so I can figure out these (infuriating/wonderful) tools.
Unfortunately I can't always share all of my work, but everything on github after perhaps 2025-06-01 is as vibe-coded as I can get it to be. (I manually review commits before they're pushed, and PRs once in a complete state, but I always feed those reviews back into the tooling, not fix them manually, unless I get completely fed up.)
I promise if someone posted human made code and said it was LLM generated, it would still be nit-picked to death. I swear 75% of developers ride around on a high horse that their style of doing things is objectively the best and everyone else is a knuckle dragger.
It wasn't something I considered at first but it makes sense if you think about text prediction models and infilling and training by reading code. The statistics of style matching what you are doing against similar things. You're not going to paint a photorealistic chunk into a hole of an impressionist painting, ya know?
So in my experience if you give it "code that avoids the common issues" that works like a style it will follow. But if you're working with a codebase that looks like it doesn't "avoid those common issues" I would expect it to follow suit and suggest code that you would expect from codebases that don't "avoid those common issues". If the input code looks like crappy code, I would expect it to statistically predict output code that looks like crappy code. And I'm not talking about formatting (formatting is for formatters), it's things like which functions and steps are used to accomplish whatever. That sort of thing. At least without some sort of specific prompting it's not going to jump streams.
Edit: one amusing thing you can do is ask Claude to predict attributes of the developers of the code and their priorities and development philosophy (i.e. ask Claude to write a README that includes these cultural things). I have a theory it gives you an idea about the overall codesmell Claude is assigning to the project.
Again I am very new to these tools and have only used claude-code because the command line interface and workflow didn't make me immediately run for the hills the way other things have. So no idea how other systems work, etc because I immediately bounced on them in the past. My use of claude-code started as an "okay fine why not give these things the young guns can't shut up about a shot on the boring shit and maybe clear out some backlog" for making chores in projects that I usually hate doing at least a little interesting but I've expanded my use significantly after gaining experience with it. But I have noticed it behave very differently in different code bases and the above is how I currently interpret that.
FWIW: “Infuriating/wonderful” is exactly how I feel about LLM copilots, too! Like you, I also use them extensively. But nothing I’ve built (yet?) has crossed the threshold into salable web services and every time someone makes the claim that they’ve primarily used AI to launch a new business with paid customers, links are curiously absent from the discussion… too bad, since they’d be great learning material too!
1. I code with LLMs (Copilot, Claude Code). Like anyone who has done so, I know a lot about where these tools are useful and where they're hopeless. They can't do it all, claims to the contrary aside.
2. I've built a couple businesses (and failed tragicomically at building a couple more). Like anyone who has done so, I know the hard parts of startups are rarely the tech itself: sales, marketing, building a team with values, actually listening to customers and responding to their needs, making forward progress in a sea of uncertainty, getting anyone to care at all... sheesh, those are hard! Last I checked, AI doesn't singlehandedly solve any of that.
Which is not to say LLMs are useless; on the contrary, used well and aimed at the right tasks, my experience is that they can be real accelerants. They've undoubtedly changed the way I approach my own new projects. But "LLMs did it all and I've got a profitable startup"... I mean, if that's true, link to it because we should all be celebrating the achievement.
Overall "meta" commands seem to work much more effectively that I expected. I'm still getting used to it and letting it run more freely lately but there's some sort of a loop you can watch as it runs where it will propose code given logic that is dumb and makes you want to stop it and intervene... but on the next step it evaluates what it just wrote and rejects for the same reason I would have rejected it and then tries something else. It's somewhat interesting to watch.
If you asked a new "I need you to write XYZ stat!" vs "We care a lot about security, maintainability and best practices. Create a project that XYZ." you would expect different product from the new hire. At least that's how I am treating it.
Basically I would give it a sort of job description. And you can even do things like pick a project you like as a model and have it write a file describing development practices used in that project. Then in the new project ask it to refer to that file as guidance and design a plan for writing the program. And then let it implement that plan. That would probably give a good scaffold, but I haven't tried. It seems like how I would approach that right now as an experiment. It's all speculation but I can see how it might work.
Maybe I'll get there and try that, but at the moment I'm just doing things I have wanted to do forever but that represented massive amounts of my time that I couldn't justify. I'm still learning to trust it and my projects are not large. Also I am not primarily a programmer (physicist who builds integrations, new workflows and tools for qc and data handling at a hospital).
It's fine, of course, to make your substantive points thoughtfully, but that is a very different kind of comment.