These days, virtually everyone leverages some work with AI. I have basically three levels at which I interact with AI, they correspond to how much I know about the process I’m trying to implement or the goal I’m trying to achieve. Coding agents get a lot of attention, but I still interact with AI a good deal via chat. The interface I choose depends on whether I want execution, understanding, or structure.

Expert Mode: Prompt and Steer

I prompt the agent and steer it until it achieves a goal. I use this when I know what is going to happen, understand the actions being performed, and can follow the underlying code. I also use it for subjects I have no interest in learning deeply. I won’t center a div manually. It’s not my job.

Learning Mode: Tell Me Why

I use chat to get concepts, general structures, and ideas before solutions. When I ask for a solution, I ask for an explanation too. I rephrase things to take my notes and copy code by hand rather than copy-paste it.

Brainstorming Mode: I Ramble, You Make Sense

I give the assistant a stream of ideas, constraints, and raw thoughts. This works well when I have a few ideas for a blog post, I’m learning a new technology, or I’m trying to adapt something to a project.

You may have recognized this pattern from Andrej Karpathy. To be honest, I’ve been doing my voice rambles for some time already; not because I’m super smart, but because I’m Latin American and long voice notes are our main way of interacting via WhatsApp. It was nice to get some validation by Karpathy himself, though.

For me, the useful distinction is not which AI tool I’m using. It’s what kind of help I need from it.