Comment by therealmarv | Hacker News Reader

Comment by therealmarv | Hacker News Reader

There are two key points which are important to get most out of LLM coding assistants:

1. Use a high quality model with big context windows via API (I recommend Openrouter). E.g. Google Gemini 2.5 Pro is one of the best which keeps constant good quality (OpenAI reasoning models can be better in problem solving but it's kinda a mixed bag). Other people swear by the Claude Sonnet models.

2. Upgrade your code tools you combine with this high quality models. Google Jules and OpenAI Codex are so brand new and have a totally different aim than Cursor. Don't use them (yet). Maybe they will get good enough in future. I would focus on established tools like aider (steepest learning curve), roo code (easier) to be paired with Openrouter and if you want to have it really easy claude code (only useful with a 100-200 USD Anthropic subscription IMHO). On average you will get better results with Aider, roo, claude code than with Cursor or Windsurf.

Btw. I think Cursor and Windsurf are great as a starter because you buy a subscription with 15-20 USD and are set. It can be most likely that the more high quality tools burn more tokens and you spent more per month but you also get better quality back in return.

Last but not least and can be applied to every coding assistant: Improve your coding prompts (be more specific in regards to files or sources), do smaller and more iterations until reaching your final result.

It's not about the tools. The complaint here is a general one that shows up no matter what tools you are using

And I tell from personal experience: The tools matter (!) for more complex problems.

Op said literally "Am I just not using the tools correctly?" and I said for the 2 of 3 tools OP tried are so new and experimental I would put them away (for now).

You're still perpetuating this myth that it is about the tools, when every interaction I've had with coding assistants and LLMs veers into the territory OP describes. To be fair, the work I am doing is pretty complex and novel.

And I use a wide range of model/versions.I mostly use Claude, from 3.5 to the newest 4.0. I also use Gemini and with Copilot.

It's not a myth. I don't promise that better tools solve every problem but it can increase quality and usefulness of the output. One of the main big problems of (nearly) all AI coding tools is to get enough context of understanding which means the right files for a prompt need to be send along. Especially the bigger a project gets the more important that is. It's kinda an AI in itself just to gather the right files for the main AI prompt which is sent to the main solving AI. E.g. in aider I specify manually which files will be sent along the prompt and they are not truncated or ignored, something which can happen easily in tools like Copilot, Cursor, even in roo code behind the scenes. This leads to the result that the main AI does not see the whole picture of the project and makes wrong guesses and maybe assumptions. More details in prompts in every tool help find the AI to understand the bigger picture and the important files. It maybe even helps to describe and save the whole project and structure in a text file to describe it to the AI better along every prompt.

Roo code is smarter with this problem, Claude Code is also pretty smart with it and in aider the problem does not exist if you as a human understand the project structure and tell aider the important context files to send to the AI.

You are perpetuating that it isn't about the tools, but it definitely is about the tools it has to do things.

The tools do matter. The original complaint is dribble. The poster does not even know what model they used in cursor which is surprising, it’s the most important part of the process. They also compared two entirely brand new tools that are not adjacent to cursor.

> The original complaint is dribble.

The usual word is "drivel" rather than "dribble".

Thanks was not paying attention with lack of sleep!

Ok, that's the "you're holding it wrong" thing.

When I start to stray into even moderately complex work, LLMs become pretty useless quick. Tell me your setup, and I will give a quick sample task that it will fail at. Stop the fanboyism please

Thank you. It is so frustrating that you hear that LLMs are a PhD level intelligence capable of any task you can throw at it and when it can't solve your problem you hear: "well you are using a1.36 and b1.37-high is really the one this time" (despite the fact that you have been hearing these claims since before that model came out) or "you are prompting it wrong, have you tried describing all of your app features and your entire approach to coding in a text file then using that to get the AI to make a list of prompts then refining those prompts into different text files and put those back into the AI..."

Who told you that LLMs are a "PHD level intelligence"?

Totally fair frustration. Unfortunately model/version does matter—it’s not pedantry, it’s debugging. And no, you shouldn’t need a prompt engineering PhD to get value, but some structure and awareness of tool limits go a long way.

It's not that I think model version doesn't matter. I switch between them all the time (often to downgrade as much as upgrade honestly). It's that I think people are misrepresenting the kinds of results you can get from these models and seem to take it as a personal attack and come up with excuses when you talk about limitations that you've encountered. It makes it difficult to engage in conversations about tools and I've gotten to the point where I don't believe anything anyone says about it anymore and I just try tools for myself.

I said people are saying the models are PhD level intelligent not that you need to be. I get a ton of value from them and I don't have a PhD.

When the original post has no clue what model they are using it throws all credibility out the window. At that point it’s appropriate to point that out to them with suggestions. Nobody here was suggesting that LLMs are PhDs like you are saying. You are the only one bringing that up.

I am more referring to my experience in general not just in the thread. I see this PhD thing a lot in the media.

> When the original post has no clue what model ...

Well, that's the point. As long as they are using a recent-ish model it really doesn't matter. Not that there are no differences in performances between models, it's that there is no model today that even comes close not requiring extensive hand-holding to accomplish real-world software engineering of even slightly moderate complexity.

Case in point: I have been frustrated that most markdown viewers don't do automatic indentation of section levels. I thought, this is a perfect test of coding assistants: the problem and solution is straightforward conceptually, and I don't even care about the platform and architecture used to accomplish it.

I've asked all the major models to implement a simple markdown viewer that could do automatic indentation, and they all fall flat. Some even give me code that will not run; of the rest, none has provided code that basically does the thing I've asked for.

I just hope you keep still trying with whatever tool you want. I see among many developers to see frustration once after some deep testing of AI coding tools and then they never look back and never try again. This means they will be stuck with soon outdated knowledge and experiences.

The LLMs advance and so does the quality. What e.g. OpenAIs o3 can do is so much better than what GPT 3.5 can do.

I'm convinced every developer who does not stay up to date with AI coding tools and know how to use them will become obsolete sooner or later.

Not knowing which model you’re using is doing it wrong, unfortunately, that matters with current-gen tools. The differences are significant. And while LLMs do hit limits fast on deep complex work, dismissing them outright misses the real utility: they’re great at the tedious stuff. No fanboyism but more middle of the road it works great for some things, ok for others and terrible for the rest.

Reply on news.ycombinator.com