Thoughts on AI-Assisted Coding

I’ve been using AI-assisted coding tools ever since the first release of GitHub Copilot Neovim plugin in 2021. So far, I’ve tried out and currently use all sorts of AI tools, such as code completion/suggestions, chatbots, and agents.

In the early days, when AI autocomplete was pretty much the only popular thing on the Internet, I didn’t follow any tutorials. Instead, I explored and adjusted how I used AI on my own, based on how happy I was with the results it provided. This became a habit, and even as more types of AI tools emerged, I’ve continued to use AI in this way:

Never try to make an AI understand the entire codebase

Letting them roam freely through the codebase and potentially exceed the context limit is not a good idea. Some might argue there are ways around this. I agree, but the point is: It’s the human who needs to understand the codebase, not the AI. When you know everything about the project, you can easily come up with a detailed, well-structured, and less error-prone prompt or plan to guide the AI agent to do the work for you without going off track.

Use AI for learning by brainstorming

AI can be a great tool to help get unstuck and generate new ideas. When talking to AI, add your own thoughts and ideas to the conversation, so that the output of AI is likelier (I think) to be relevant and useful. You can try asking yourself dumb questions and answering them¹¹ I learnt this from a blog post of Terence Tao, famous mathematician and Fields Medalist. See Ask yourself dumb questions – and answer them!, then brainstorm with AI to get more ideas.

Let AI do the boring stuff that does not teach you anything new

Humans are lazy by nature, and that is not necessarily a bad thing. Use AI to handle the tedious tasks that teach you nothing new, such as adding debug prints, writing tests for trivial code, or performing simple refactors. That way, you can save your time for more creative and critical work, such as designing the architecture and APIs.

Double-check the output of AI

It is your responsibility to verify the results, as only you possess good taste, good sense, and the ability to define the gestalt of the project. You don’t? Then use AI for learning, and improve yourself so that you can do so in the future.

See also Mario Zechner’s²² Author of pi blog post about how we should and should not work with agents.