It’s a pretty unusual feeling diving into the deep end of a project, not really knowing where you are going or what lies beneath. It’s even scarier when you are not sure if you’re in a raft or a yacht before embarking.
When I made the stubborn decision to build my own system for Gravity DesignWorks, Inc., I only knew what I knew: above-average confidence in WordPress, enough coding knowledge to break things, and that using AI tools to get the first part working had already been a success.
My initial attempts were with a free version of Claude. It became evident very quickly that the usage limits were going to hinder my progress, so I opted for the Pro plan, hoping that was enough to get me rolling.
I was wrong.
I quickly ran out of usage on my first day, and then I was also out of credits for the week.
Hard stop.
I had been using ChatGPT for quite a while to help define project scopes, research ideas, and put together more professional and robust proposals than I could produce manually under tight timelines. So I decided to seed it with the progress I had made with Claude and see if I could pick up where Claude had left me stranded.
My first observation while piloting the ship with both Claude and ChatGPT was that these were two very different tools that worked in very different ways to reach similar outcomes.
Switching to ChatGPT allowed me to run through iterations and revisions much faster, whereas Claude was much more thorough. I quickly realized that using Claude for iterative building was really just burning through usage limits. I could beta-test new ideas in ChatGPT and get to testing much faster.
With Claude in my back pocket as a second tool, I was able to run fast prototyping in one system and then check architecture, schema, and more robust code in the other.
While this was extremely useful and productive, it also carried some flaws.
You start to get scope creep between systems.
You get well-fleshed-out code in one app and quick-and-dirty working code in the other.
Using both tools together ended up creating a mix of coding styles across the project.
But the bigger realization came a little later.
At some point it dawned on me that I had become the API linking the two systems together.
I was the translator between them.
And figuring out what context to give each tool turned out to be its own learning curve. Sometimes you dump the whole file in. Sometimes you summarize. Sometimes you isolate a bug. Other times you have to step back and talk about the bigger architecture.
The AI knows a lot about programming. It knows nothing about your specific system unless you explain it properly.
No matter how good the AI tool is, or what level of access you pay for, you as the system architect still need a deep understanding of the subject matter. You have to know when the tools are wrong and when you need to take a harder stance steering the ship.
You can’t leave the Falcon in autopilot while the Wookie pretends to drive.
AI didn’t build the system.
It just made it possible for one stubborn member of the coveted 18–49 demographic to take on the workload that would normally require a medium-sized team.