An AI can produce a long argument while leaving its reader unsure which assumptions each step needs. In this first-person essay, Dynomight describes the less glamorous part of doing mathematics with current models: breaking tangled arguments into smaller pieces, improving notation and checking how the pieces connect.

Sometimes that clean-up reveals a stronger result. The account also describes losing hours to almost-working answers and the temptation to send collaborators material one has not understood.

A useful workflow reflection for anyone using AI on difficult problems: ask for self-contained steps and make their assumptions explicit.