AI Driven Development

Published See discussion on Twitter

Over the past week or so I’ve taken some time off work (for a friends wedding) and found myself with a bit of free time - so I figured I’d work on a side project. The idea of project was still pretty fresh, and I figured that I’d try to lean into using ChatGPT to help me build it out.

Before digging in, I have been thinking more and more about AI as of late, I’m still really new to whats available and how it all works, but I’ve been trying to use the available integrations where I can. I pay for the upgraded version of Notion to use their AI integration (which I honestly haven’t really used all that much yet), I also pay for the upgraded version of Raycast with their AI version (which I’ve so far found to be really direct and straight to the point - I’ll talk a bit more about this later on in the post), and mainly I’ve been using ChatGPT.

I’ve formed some rough thoughts that I wanted to try and put them all into a post on this blog - I may revisit some of these in the future, but we’ll see!

Thoughts on AI in general:

It’s pretty decent, but I feel like I haven’t found a really solid fit for how it can wedge into my workflow. I don’t often need to write creative prose pieces which feels like where it really excels at the moment. However I do want to continue to give it a lot of thought and experimentation. There have definitely been many times where I’ve thought that different workflows that I do for work could be made more efficient, but it feels like no one is exploring them yet (at least not publicly).

A core example that comes to mind, and is also top of mind for me at the moment is an AI backlog assistant.

I want something that crawls our backlog, triages inbound requests, helps us to answer why we approached a problem in a specific way when looking back on it, and also helps us to better understand how we can tackle work.

I feel like this would greatly benefit from being integrated into not only our project tracker, but also Slack and GitHub where we have most conversations and where most of our work happens.

I’m sure I’ll have some more thoughts about AI use cases in general, but for now I want to dig into the focus of this post, which is how it works for web development.

Thoughts on AI for web development:

I’ve tried GitHub Copilot in the past for about a month or so (basically the free month that you can get from GitHub), and at the time I felt like it didn’t really fit into my workflow all that nicely. It felt like a slower version of the features/benefits that I got out of the TypeScript Language Server in VSCode[1].

This past week or so however I dug into a new project using Go lang and SQLite leveraging only ChatGPT (v3.5, I haven’t purchased the 4.0 variant yet). Overall, I enjoyed the setup quite a bit - it was nice to be able to talk to ~~someone~~ something about the problem that I wanted to solve, and have it work with me to write the code and implement changes.

There was definitely some friction - at best it felt like pairing with a junior engineer, but not a great pairing session in that. I often needed to ask the question two or three times for it to actually do what I wanted. Only about 4 or 5 times during the work of the project did I feel like I hit a wall and I just needed to implement the changes that I wanted to do manually.

Another point of friction for me was needing to go back and forth between my editor and my ChatGPT tab, including several times of pasting in the error / code I had to keep it up to date on what the code looked like after I made some changes. I think that GitHub Copilot now offers a chat like feature that knows about the state of the codebase, so i think that would improve the iteration cycles a bit more also.

In Summary

Overall, the experience was pretty neat - I definitely want to lean on ChatGPT more for working on more greenfield projects. I think the place where it really excelled was that I was fresh to the project as well - so I could discover the problem space and the eventual solutions as well while working with it. I feel like it would be relatively complex to dive into an existing large project and have it up and running and providing valuable feedback along the way. Maybe when it has a deeper integration into the codebase and the history of the project then it would be able to be a reliable assistant.

[1] -

I do wonder if that opinion would have changed had I start using it in a language that I don’t know as well, such as Go lang.