Earlier this year I gave the same sixty-minute talk enough times to notice that the useful part always started right as the clock ran out. So I made the long version.
It’s called A Short Book on Context Fundamentals. About two hours to read, and free: EPUB, or the markdown source if you’d rather read it on GitHub.
The book is an argument for the idea that drift is a context problem, not a model problem. When a session goes sideways an hour in, the model didn’t get dumber. What it can see changed, usually without you noticing.
The rest of the book is the practice that follows from that. How to write rules that survive across sessions (short, invariants only, no motivational speeches). How to shape a command so the model doesn’t have to guess. How to hand it knowledge just in time instead of just in case, and how to carry all of it with you across tools. There’s a tour of the 2026 tools, which will age fastest, and a chapter on the role I think teams are going to need soon: the person who owns the context.
It’s for anyone who uses these tools seriously, not just engineers. I ran the workshop for designers and PMs as much as developers, and every chapter ends with something to try on Monday.
Books like this are usually a slog and I didn’t want to make another one. So it reads more like Bill Bryson than a manual. The model is a brilliant new hire who walks in fresh every single morning, the context window is the packet of papers you hand them at the desk, and the book is about what goes in the packet. There are asides and a few jokes. I had fun making it, partly because it was fun to read back, and I think that’s the only reason a book about a moving target survives more than a few months.
I finished it in April and then told approximately no one, which is a habit I’m working on. The handful of people who found it anyway said it helped. If you read it and something is wrong, or useful, tell me.
(It’s built with Bookbind, a small EPUB builder I wrote so I could bind a folder of markdown into a book.)