AI changed what I could want

For a while I thought AI had simply removed friction from work I already wanted to do. Then we built a small tool that let an agent read Figma as text. We were trying to save tokens. Once it existed, I started asking what it meant to design for a reader that wasn’t a person. Soon I was writing specs and even notes to myself differently. I only began wanting to work that way after the reader existed. I don’t know how often that happens. Here, I could see it happen.

That change reached the team in a leadership review. We brought a few working prototypes for people to try. Usually I would walk through one direction and describe how it was meant to work. This time people could use several versions and compare how each one behaved. They found weak points in the work, and the questions in the room came from what they had just tried. After that, talking through one direction no longer felt like enough.

Trying the prototypes gave us better evidence. We still had to choose which direction to keep. The choice depended on my judgement, and I was also using the model to test it. Its default was to agree with me, so I added “push back if you disagree” to its instructions. At first, having to ask felt reassuring because I was the one asking to be challenged. After a while, I started to expect the challenge. The tool had made disagreement part of what I wanted from it.

So I am changing how the team works while the tools are changing what I notice and ask for. I can remove an instruction tomorrow. It is harder to unwind a review format after the team has begun using it. A useful experiment can become normal before I know whether the judgement behind it will hold. We still have to decide which experiments to keep. We should be able to change our minds after a useful one becomes familiar. I don’t yet know what would let us do that. It is quite fun trying to figure it out.