Macfadyen splits the problem into three stages: what goes in, what happens while the system thinks, and what comes back. Each stage has its own failure modes, and her argument is that most failures blamed on the model are interface design failures. She opens with the ELIZA effect, from 1966, to make clear that attributing understanding to a machine that writes well is not a novelty of this decade.
What I took from it
Show the context. Before the user types anything, the system has already assembled its own package: history, the open document, preferences, the output of earlier tool calls. The blank box implies the opposite and produces answers that look arbitrary. Since then, in anything I deliver that answers over documents, the screen has to say where the answer came from.
The three channels through which intent arrives: the context the system gathers by itself, what the person writes, and what the person selects or points at. When an AI feature fails, the failure can almost always be placed in one of the three, which turns a vague complaint into a specific fix.
And the point that waiting communicates. A spinner that turns for eight seconds without saying anything is saying something, and it is rarely what you intended.
What I left out
The assumption that there is a product team with a designer behind all this, with user research and test cycles. There is also a tension the book never resolves: it announces that principles age better than patterns, and then spends many pages on patterns tied to this year's products. The principles are the part that earns the price.