Build Wide, Ship Narrow: How AI Changed the Economics of Code Review

The old approach to software development required engineers to decide boundaries before they started building. Developers would write an RFC describing the feature, split it into smaller issues, and build them sequentially. This structured workflow kept code reviews manageable and avoided big-bang merges, but it also asked engineers to make critical structural decisions without knowing much about the problem at hand.

Sometimes these assumptions turned out to be wrong, leading to wasted effort and the need to discard previous work.

However, the economics of code review have shifted dramatically in recent years. Three major factors have become dramatically cheaper: building code using AI assistants, designing plans that can be reshaped at conversation speed, and splitting finished branches into small PRs. These advances have made it possible to build wide and ship narrow, a workflow that combines building functional code with early product validation.

In this new approach, the design process still comes first. Before touching the editor, a plan is created through adversarial rounds that challenge the idea. Once the plan is solidified, the design is committed as a spec, even before any code is written. This spec serves as a foundation for building wide, committing save points as you go, and demoing the feature before anyone reads a line of code.

The key change is that after the design is settled, engineers build the feature end-to-end without stopping to open PRs and wait for review. Commits serve as save points, proving concepts or indicating risky experimental changes. This wide-building approach allows for rapid iteration and real-time feedback. Once the feature works end-to-end, it is demoed to stakeholders, avoiding lengthy review cycles on a flawed direction.

Finally, the AI-powered splitting of the finished branch into a sequence of small, independent PRs enables narrow shipping. These PRs are cut fresh from the main branch, each representing a clean, reviewable change. The natural seams in the code reveal the appropriate boundaries, eliminating the need for pre-decided boundaries. The result is a review process that focuses on narrow, understandable changes, maximizing the value of human code review while minimizing the time spent on narrow PRs.