I don't know if he wrote anything formal, but I remember from talking with him, from his critiques of particular interfaces, and from projects that he was interested in, that he favored UI that made operations as concrete and manifest as possible, that made it as easy as possible to discover operations by experimenting, that made mistakes as low-cost and painless to recover from as possible, and that featured direct manipulation.
The basic idea is that modes make the same action (pressing the "D" key, for example) do different things. They make things easier for programmers who want many operations on machines that only have a few possible actions, but they make things hard for the user who have to pay attention to the current mode and know how to navigate from that to the mode where the desired operation is possible.
Fascinating because that exploratory process is essentially what we recognize as "graphical user interface" as of today, and the whole industry has committed itself to that particular design, to the point that exploring how to build interfaces from any other set of principles feels like a titanic task.
Is it a blanket rule for him for all interfaces, or just text editors?