Hacker News Reader

Comment by skeeter2020 | Hacker News Reader

skeeter2020

You can't possibly believe this, or you and me (and many others) are doing something different. LLMs have created an entire new - huge - set of bang-your-head moments, as they go off half-cocked in a million simultaneous directions, chasing their tail, or just making shit up. And since the vast majority of work is on existing - often ancient - codebases, let's find out if you feel the same way in 18 months.

View on HN

LLMs are great for anyone who isn't responsible for the consequences of what they code.

Maybe I'm weird, but my usage has been very conservative. As in, I treat the LLM like a junior dev that I have to micromanage and handhold.

I am terrified of allowing these things to complete tasks end-to-end with nothing intervening. Maybe that's why I don't run into many of these issues. I mostly delegate grunt work and manual tedium, not reasoning or design choices to the LLM. I may consult the LLM and ask for criticism, but there is no way I'm going to allow it to quietly make design decisions that I don't know about.

That's only if you do agentic coding.

I use LLMs in the following ways:

1. Copy-pasting code into the web chat UI and asking for something (bugfix, add a feature, refactor, explain, review it etc), including entire source code files. A $20/mo Gemini subscription goes a long way (never been rate-limited). I only use the highest model. I often just copy-paste the entire source file between 3 backticks.

2. Cursor Tab. I do have hotkeys to enable and disable it; it's disabled most of the time otherwise it gets annoying.

3. Single-file changes directly from Cursor's AI sidebar. I only do this for simple, predictable stuff because even their auto-routing "Premium" setting is not as good as pasting stuff into Gemini 3.1 Pro.

That means I have only two $20/mo subscriptions: Gemini and Cursor.

I don't use Claude Code, it's really for people who don't know how to code. I don't use Plan Mode; I make and track the plan myself (if at all). I only tell the LLM granular tasks to execute. I don't use `claude.md` or `agents.md` or anything like that. If I don't like a particular output, I reset everything, modify my prompt and try again.

I believe this is the only way to fully leverage LLMs without losing any product quality. If you're trading off quality for "speed" (in quotes because over the long term, a low quality codebase is a massive drag on productivity) then there's no point.

I _think_ what you’ve said is “go shallow, not deep”. That is, don’t let the walk you make inside the latent space a long one. Twenty-five short and peppered steps, from de novo, is better than one long, protracted stew.

Is that accurate?

Well yeah. If you know what you're doing, why would you let the AI take control?

Well, if it works on step one, then why not step two? Where would different folks draw the line? My grandparents might continue on a while, whereas I would not. But if it also “works” on step two for me, should I take a third?

What counts as “works” is the important bit, I think.

Yes, if you're using them to write large chunks of code or entire features. If you just use them to clear up some trivial problem in an unfamiliar technology that you used to spend 30 minutes googling with 50 tabs open, or stuff like write a method to filter, map and reduce an array based on specific criteria, they're a godsend.

Give them work in smaller chunks.

You are in charge of what the LLM does. If it's running off half-cocked in a million simultaneous directions, that's on you. Write better skills. Tell it not to do that. Break into its loop and ask it wtf it thinks it's doing. If it's making shit up, force it to test more.

The LLM will do what you tell it to do. Manage it.

Reply on news.ycombinator.com