Using AI to become a one-person UX lab
The divided-attention problem
I had used AI to plan work, write software, and investigate technical failures. This was the first time I asked it to watch me use something I had built.
Testing your own product creates divided attention. I needed to use Spork naturally while tracking time, recording confusion, remembering my expectations, and investigating the system when something felt wrong.
The more attention I gave to documentation, the less attention I could give to the experience. I needed another person in the room, so I used AI.
Spork is a digital office where AI agents plan and perform software work. I wanted to test the complete experience against a real task while staying anchored to the original product direction.
Putting AI in the observation room
I gave an assessment agent the product vision, relevant requirements, and an earlier scorecard. It helped turn them into a test protocol with tasks, questions, timing checkpoints, and technical evidence to collect.
During the session, the agent told me what to attempt next and asked what I understood from the screen. It recorded my expectations, hesitation, and confusion while I concentrated on using the product.
It simultaneously watched logs, subprocesses, files, and verification records.
This produced two synchronized accounts: what the system was doing and what I could tell it was doing.
What the two views revealed
The test objective was to build a command-line changelog tool. Spork produced a five-milestone, eight-ticket plan and dispatched eight AI workers into an empty sandbox. They created a functioning npm package with twelve passing tests.
The assessment agent asked what each worker was doing, which ticket might finish first, what file was changing, and whether anything was blocked. The office floor gave me no useful answers.
At one point, the screen produced roughly ten minutes of dead air while the logs showed two workers actively running.
The interface view captured the inactivity. The technical evidence proved that work was happening. Connecting them exposed the real problem: the system possessed useful information that the interface failed to communicate.
Making observation part of the build
AI helped me create a small observation lab around a real human session.
I remained responsible for pressing the buttons, interpreting the interface, and describing where I felt confident or lost. The agent handled the protocol, timing, notes, technical inspection, and comparison with the product goals.
Testing my own work does not replace watching other people use it. It gives me a disciplined way to find problems before asking anyone else to spend time with what I am building.
I could stay inside the experience while the agent watched the product from both sides. AI made that level of observation practical enough to become part of the build process.