Shift review from code to design when AI writes the code

It positions itself against the current "merge without reading" debate, which mostly argues over which parts of code review people should read and which to hand to automated checks or AI review. The author instead moves the question one step earlier, to the plan the AI writes before any code exists. In his team's vocabulary, the plan is the document the AI produces before implementation, the design is what that plan decides (what to touch, the shape of data and APIs), and the instruction document is what implementation is handed once a person has checked the plan.

His case for the shift rests on volume and cost. More work delegated to AI means more pull requests in front of people, and reading all AI-written code is described as impractical on volume alone. Design errors also get more expensive the later they are found; if the shape of the data is wrong, much of the written code has to be redone. A plan is shorter than code and states intent directly, so it is faster to read than hundreds of lines of diff and easier to spot mistakes in, and because nothing has been written yet, a fix costs only the plan.

Implementation then follows the approved plan. The implementing AI is not allowed to make design judgments, and AI review checks whether the code matches the instructions, with a light review run on each implementation and a heavier review run in parallel before the PR is created. The author concedes that design review adds human effort, but says it is worth it, and he also mentions his personal macOS app tanacode, which he says is not an official Anthropic product, as a way to inspect changes before opening a PR.