Writing
Writing
Notes from building an autonomous assistant, a design system, and the machinery to tell whether either of them still works.
- Forty-Seven Palettes I gave my assistant fifty skins and shipped thirty-nine I couldn't actually read. Why the generator produced them, why my first fix passed every contrast check and still looked worse, and what a design system written for agents did for the repair.
- Fourteen Empty Strings My eval scored a prompt at zero across the board. The prompt was fine. The eval was measuring nothing. Four lessons on building evals that can actually fail, and an audit that found eleven gaps in my own.