Ask HN: How to Start with LLMs/Vibecoding? | Hacker News Reader

Ask HN: How to Start with LLMs/Vibecoding?

1 pointe-topy8 comments

Inspired by LLMs lately becoming (terrifyingly) capable, and my new work project being no longer fun, I wanted to ask: how should I (or anyone experienced in programming but new to using AI) get started with AI in programming?

I understand a lot of the theoretical basics (NNs, LLMs, Agents), but I haven't really tried using them for my professional work. I'd normally just search the internet, but considering the amount of conmen and slop, I'd rather ask people who actually use it. Guides, blog posts, videos, anything useful or informative would be very appreciated.

For context, I am studying CS and have been working as a full stack ~junior for 3 years now.

View on HN

You will have to decide how you like to work, some people are trying to go all in on how agents work: let them run long sessions, set constraints through /skills, evals and context but then trust them to fully implement the code (dont even bother reading any code, reading code is an anti pattern). Others use AI like a junior developer, giving clear discrete tasks, monitoring and steering for architectural and coding style course correction and working on teaching it better practices through context management and rules. I am in the treat it like a junior camp but have also had products where I end up reading less code than I would expect.

You'll notice the tools push different perspectives, coding in VS will probably push more of the discrete task, AI as a junior dev approach, claude code/codex/pi/opencode etc can be more "autonomous, dont even show the diff" vibes with long running tasks that spit out 10,000 lines of code without even asking for any feedback or review.

What I have found is that planning documents are critical, start with an agile roadmap that links to epics files outlining large features. They should have clear acceptance criteria that act as gates to close out stories/sprints/epics. Every document should say "follow red/green TDD" so that testing discipline is a firm requirement.

I have also found that AI will go rogue no matter how good your context, planning, evals or TDD, so always work on a clean git branch so you can blow it up if the AI does something wrong. Treat AI sessions as experiments and prototypes and be careful of feeling sunk cost fallacy. Code gen is cheap so do implementations 4 times until you get "good" code/ui/ux architecture etc.

I would be happy to do a pair programming session to show how I do things. I built www.propelcode.app an agentic IDE so I can experiment with ways of coding and not be dependent on the frontier labs or big company products.

https://calendly.com/ryanwmartin/open-office-hours

If your aim is to just get started, get yourself a Claude account and just download Claude Code. You don't really have to read anything on it, as you use agents you will get to learn what they are good and bad at but defaults are quite good anyway if you use a capable model like Opus 5. Truth is there are a million guides and plugins in the open but they really provide incremental gains for a beginner as they would overwhelm you.

If you are more conscious of the price, try out opencode or pi and research OpenRouter or similar for coding with models like GLM 5.2 or Kimi K3 or even Deepseek 4, they will be less out of the box but cheaper.

I have been using OpenRouter for a while now, and my Uni is self hosting Kimi K2.7 and Qwen 3.5, so I do know my way around chat completion. I've tried Claude Code with Kimi but it just felt confusing to use (probably cause I didn't know what I was doing) and the results were bad (probably because the codebase I've inherited is a mess, and it can't comprehend the web server logs actively lying (and neither can I)).

I've heard Claude Code alternatives can be better for non-claude models, but again, which ones? How can I tell if a tool is good or bad when everyone has the same LLM'd readme? How can I avoid security nightmares like OpenCode when they seem so popular?

How card can it be? You have a favorite IDE, right? Pick an AI service you can afford. If you are a student at an institution you probably get heavy discounts, if not pick the cheapest monthly tier on an AI service you respect. If you are not sure about committing to a subscription then just buy tokens on Deepseek.

Find an agent plugin aka a harness for your IDE that doesn't require a subscription, unless of course you enjoy giving money away. Some agents are configured to work with particular providers, others are 'bring your own key', so again it depends on whether you like being locked in to someone else's ecosystem and paying acordingly. you can also get harnesses that are self-contained rather than being part of an IDE, but I assume you like your existing tools.

Take the API key you got from your AI service, plug it into the agent. Start building a toy project; the easiest is something that you have previously made yourself so you know something about the data structures involved, problems you're likely to run into and so on. Treat the agent like your assistant. I prefer asking it to draft a plan of what it's going to do which I then approve rather than just setting it free and then puzzling over the output, you may evolve a different approach.

Well I've done everything you're saying, but I got stuck after that. I've tried claude code, I've been using OpenRouter, but after that it just seems like a million different options.

Take tools, there's a million different options, and I don't know how to choose between them. Or the general mentality of treating it as an assistant vs the pilot. Or the massive ecosystem of skills and MCP. It just feels like a massive unexplored ocean to me. But I guess I'm just getting analysis paralysis.

None of that is necessary. It’s like people who install Obsidian and then spend weeks obsessing over the “top ten must-have plugins” for the ecosystem instead of just using it as it’s intended: as a tool to take notes.

The purpose of a coding LLM is basically to assist with building something. So start with an actual concrete goal: something you want to build. Maybe even take a past project and try to recreate it, but this time using an LLM-augmented workflow.

Don’t get bogged down in a bunch of orchestration/MCP/skills/etc. - keep it simple and expand as you go. The best thing you can do is actually use these tools to help you develop an intuitive sense of their strengths and weaknesses.

You don’t need to watch videos or tutorials. Grab something a basic agent harness like oh-my-pi [1], throw $30–50 on OpenRouter, and start playing with it.

[1] - https://github.com/can1357/oh-my-pi

You are. Obsessing over tooling is a form of procrastination. Once you have it up and running and it can generate code, documentation etc and write and save files, just start building things. you can try one-shotting it (Clause, build me a better version of Thing), or tell it what you want to do and have it sketch the framework out for you or whatever. You might find it easier to work with long detailed specifications, or to limit your prompts to a paragraph at a time and keep laser focused. different models have different benefits and they also have their own pain points, like getting carried away and writing too much or starting to loop back to earlier in the context window etc. etc. Just roll up your sleeves and start playing. You don't know what you want to optimize for yet so you're just comparing yourself to the fake personas of other people on the internet who are trying to pose as the next Elon Musk. After you have been doing it for a while you can refine your own workflow. But focus on outputs more than workflow.

You have a huge advantage from your CS studies in that you're already knowledgeable about algorithms and data structures and the concept of algorithmic efficiency, so once you get prototypes of your ideas up and running, refining them should yield rapid improvements.

Build Applications with Local AI Models on a Mac

https://www.manning.com/books/build-applications-with-local-...

Reply on news.ycombinator.com