Member-only story
We Blamed the Model. The Prompt Was Contradicting Itself.
How a pile of prompt patches hid a design problem, and the three-day tool that made my team see it
My team was shocked by what they saw on the screen. A dense tangle of nodes. Next to them, a wall of prompts that contradicted each other. They had built every piece of it.
I had put that view together in about three days, in between other work. Here is why I built it, and what it showed.
The setup
I manage a team that builds a service on top of AI agents. It generates written content that businesses publish under their own name. I joined after launch.
By the time I arrived, the prompts had grown enormous and the output was unstable. The team blamed the model’s limitations. The real problem was the design.
The complaints
Most users publish what the service writes without touching it. Some edit it every time, and those are the users who send complaints:
- The phrasing isn’t to their liking.
- The tone doesn’t fit.
- The service name is slightly off.
- The text says something that isn’t true.
