Without an explicit concept document, each AI prompt produces a locally competent component. The components do not share architectural assumptions, because each was generated independently. Local optima accumulate. The resulting system has code that works but lacks conceptual integrity — and without conceptual integrity, modification becomes a source of surprise rather than a predictable operation.
The AI-era architect's notebook therefore has two functions. First, it disciplines the builder: forcing her to articulate what the system should be before generating code, which is harder than it sounds because the AI makes building so easy that starting without a concept feels cheap. Second, it disciplines the AI: giving it the persistent context that its architecture does not provide, so that each prompt can be evaluated and, when necessary, constrained by the project's established principles.
The practice has a measurable output: a document that the builder returns to repeatedly, edits as the project evolves, and uses to evaluate AI-generated components before accepting them. The discipline is demanding. The payoff is that the system, however fast it is built, remains its own thing rather than becoming an accumulation of locally optimal fragments whose collisions only become visible under maintenance pressure.
Brooks would have recognized the AI-era notebook as the same instrument he advocated in 1975, operating under new constraints. In 1975, the notebook constrained the team. In 2026, it constrains the builder and the machine together. The underlying insight is unchanged: a system that does not have a concept cannot be extended without losing its character, and the concept must be held explicitly because the pressures against it are constant.
Brooks introduced the notebook concept in The Mythical Man-Month (1975), drawing on his OS/360 experience and on the architectural-documentation traditions of building construction. He treated it as a mechanism for propagating design intent across a team; its AI-era extension generalizes the mechanism to propagate design intent across the builder and the tool.
The notebook records principles, not just decisions. Principles generalize; decisions do not. The notebook's value is in its capacity to answer questions the architect did not anticipate.
Without a concept document, AI-generated code accumulates local optima. The components work; the system does not cohere.
The notebook disciplines both builder and machine. It forces the builder to articulate the concept; it provides the AI with persistent context its architecture lacks.
The notebook is a living document. It evolves as the project evolves; it is not a one-time specification.
The discipline is hard because AI makes its alternative cheap. Building without a concept feels faster; the cost is deferred and paid at maintenance time.