Why I Started Building Software with AI Coding Agents
I started a few months ago to build my own platforms. Here is what has worked, what has not, and how I keep the results under control.
A few months ago I started building my own software with an AI coding agent. I wanted to learn better ways to build my own platforms and systems, and this was the fastest way I could find to do it.
My tools
I use Claude Code for all of my coding. For everything else I use Codex and ChatGPT: image creation, creative text and day-to-day questions.
What I have been building
My projects include this website, a community website and Discord bots for Goated Geese, and a Shopify theme for a streetwear brand I am developing. I have also worked on a community hub for the game WARDOGS.
What went well
This website came out clean and has worked well.
What did not
The WARDOGS site has been messy and hard going. The main reason is not the code. It is that I have had trouble getting access to the data from the systems it needs. No coding agent can build around data it cannot reach.
How I keep it under control
I use OpenGSD for almost all of my work. It uses a lot more tokens, but it sets out the specification clearly before anything gets built. I wrote about how that works in how I keep AI-built projects under control.
Trust, with checks
For now I trust the output as long as I use my own brain alongside it. I check things constantly, and I prompt the agent to check its own work and to secure things properly.
I also think AI-written code tends to bloat. It may need simplifying later, and I would rather say that now than pretend otherwise.
Limits
I am on Claude's 20x plan, and with a lot of projects running at once I still hit the usage limits. Time is the other limit. Both slow down how quickly my projects move.
Comments