How to Prompt (Communicate) Better?
Why better prompting is really better communication—and what AI agents need from us to succeed.
It’s not uncommon to see people post screenshots on their AI chat histories when AI agents grab something. People angrily shout at the AI agent about “Why did you do that thing wrong? I’ve told you, blah blah blah. Why did you steal, blah blah blah?” these kinds of things. I understand the frustration when people find their AI agent didn’t really follow what they wanted their AI agent to do but you cannot think that it’s all their agent’s fault. It’s not all their fault.
Ever since Claude launched its 3.5 models, I’ve told myself: if I give an AI agent a task and it doesn’t perform as I wanted, that’s not its fault. It’s mine. I didn’t give it clear instructions on what I wanted. Now frontier models are very intelligent now, and AI agents are very competent. I think it’s increasingly a skill issue on the human’s side when an AI agent doesn’t perform as well as expected.
I’ve found that the best prompting techniques are really just communication skills. Think about asking a colleague for help at work: you’d give them the relevant documents or email chain, based on what you know they already know and what they don’t. They don’t know about a support case unless you tell them. But once they have the system knowledge and the details of the user’s inquiry, they can resolve it well.
That’s the same for AI agents. You need to know:
- what AI agents can do
- what AI’s knowledge is for
- what tools AI could use
- what things you need to provide to that AI
- what things you do
Then you can let the AI do the things you want at a higher success rate. In the end, prompting well is just communicating well.