You’re Probably Using AI Reasoning Models Wrong

You might have tried one of those new AI reasoning models, maybe O1-Pro, and thought "this thing just doesn't get it." I've been there. As a software developer who uses AI for coding daily, I want to share something that completely changed how I work with these models.

The Claude Way

Most of us are used to working with Claude 3.5 Sonnet (or GPT-4o). Claude is great for coding. You tell it what you want, give it some context, and you get your code. There's often bugs or other shortcomings, but working with Claude is always straightforward.

I've gotten pretty good at being selective with context. You learn to give it just enough information - not too much, not too little.

The Reasoning Model Way

Then I tried out reasoning models like O1-Pro, and hated it. I used it exactly like I used Claude, and guess what? It sucked. Instead of clean, simple code, I got these massive walls of text with endless explanations. It felt like reading a research paper when all I wanted was some code.

But here's the thing: I was using it completely wrong.

There's two important things that I was missing:

1. Reasoning Models Are Context Hungry

Unlike Claude, where you need to be picky with context, reasoning models want ALL the context. Seriously, dump as much relevant information as you can on them. They're actually really good at figuring out what matters and what doesn't.

2. You Need to Spell Everything Out

This was the real game-changer. While Claude is great at figuring out what you want, reasoning models need explicit instructions. Want code without the essay-length explanation? Say it. Need a specific format? Tell it. Want to know which files to put the code in? Ask for it directly.

How I Use Them Now

These days, I use both types of models, but for different things:

• Claude: This is still my go-to for quick, everyday coding tasks. It's fast, intuitive, and just gets the job done.
• Reasoning models: I pull these out for the complex stuff. When I need to think through bigger architectural decisions or when I'm dealing with tricky system-wide changes.

Making Reasoning Models Work

1. Don't hold back on context - give them everything you've got
2. Be super specific about what you want

Once you understand how to work with reasoning models, they're incredibly powerful. It's like having a senior developer who really thinks through problems before proposing solutions.

So if you think reasoning models aren't working for you, you might just be using them wrong. Give them more context, be explicit about what you want, and watch how they transform your workflow.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com