How can I improve my communication with AI?

To be honest, chatting with AI is a love/hate thing. Obviously, I think I have said very clearly, but it gives you a "wrong answer". Recently I realized that communicating with AI is like teaching children, you have to pay attention to the method. Take writing code for example, if you just say "there is something wrong with this code", the AI will give you an ambiguous answer. But if you can treat a new colleague like, the background of the problem, the error message, the expected results are clear, the effect will be very different.

Think of AI as your new coworker

I have a friend who works as a CTO at a startup, and their team recently came up with a particularly powerful AI assistant. At first everyone used it as a search engine, with predictable results. Later they found that if the AI as a new member of the team, like a newcomer to patiently explain the background of the project, the technology stack, and even the business logic are clear, the AI gives a program quality straight up. This reminds me of last week to help interns debug experience - you have to let him understand the whole process before he can find the problem.

It's so specific it's sickening.

Once I wanted AI to help me optimize a sorting algorithm, the first said "optimize this function", the result of the program it gave is simply impossible to see. Later, I learned my lesson, the input data size, performance bottlenecks, and even business scenarios, special requirements are listed, such as "this function to process 100,000 user data, the current time consumed 2 seconds, the requirements of optimization to 500 milliseconds or less, and the memory occupation can not exceed 50MB". Guess what, AI actually gave a wonderful solution based on partitioning, better than I expected.

This reminds me of the questions that get a lot of likes on Stack Overflow - they all have one thing in common: the problem description is ridiculously specific. For example, instead of asking "Why is my Python code giving me an error?", you might say "In Python 3.8, I encountered a ValueError:... when processing a CSV file using pandas 1.3.0. The following code snippet can be reproduced...". It seems that communicating with AI has to follow this golden rule as well.

The Art of Iteration

What surprised me most was that talking to the AI was actually an iterative process. Not satisfied with your first answer? Don't give up in a hurry. Just like when we do product requirement reviews, the first version of the program often needs to be polished over and over again. I have a friend who does data analysis, he keeps a complete record of the conversation every time he discusses with AI, and cuts back and forth like a ping pong game. He says this is much more efficient than asking a perfect question at once - after all, even humans often need to double-check their communication with each other, let alone AI.

Recently GitHub Copilot came out with a new feature that allows you to ask questions and make adjustments to AI-generated code, which takes the idea of "iterative communication" to the extreme. I tried it out, it's like having a programming partner sitting next to you, you say "this variable name is not intuitive enough", and it immediately gives you five or six suggestions for improvement. This kind of incremental dialog makes the problem solving process very natural.

Give the AI a role.

Another particularly interesting trick is to give the AI a role. For example, if I want to write a technical document, I will say "You are a Python technical writer with 10 years of experience, please explain this algorithm in layman's terms". The AI automatically adjusts the tone and depth of the content, giving a clear structure, even the examples are very relevant. This is also useful when reviewing code, letting the AI play the role of a "strict technical supervisor", its nitpicking is more serious than a real supervisor.

However, it is important to note that the characterization should not be too wild. I once asked an AI to "play Steve Jobs to review my product design" and got a bunch of empty words filled with "reality distorting force fields". Later, I realized that specific professional roles like "senior React developer" or "strict security engineer" work best.

At the end of the day, the key to communicating efficiently with AI is to transplant those good communication habits from our regular jobs. Clear context, specific requests, patient iteration, and a little role-playing creativity. Of course, AI isn't a real person, after all, and it can't read minds - but if you're willing to spend an extra 30 seconds to make the problem clear, the payoff could be saving three hours of debugging time. That's a good deal, any way you look at it.