Rethinking Software Development Workflow in the Age of Generative AI

Generative AI has made it significantly easier to produce large volumes of code. However, this code is often more verbose than human-written code, appearing bloated and handling edge cases that human developers might not naturally consider.

While this can sometimes enable developers to build functionality beyond their traditional limits, it more often results in bloated software and reduced overall code quality.

As programming shifts from using formal programming languages to using natural language, the generated code becomes an intermediate representation — and the true “source code” is the English description of the task. Consequently, the quality of the specification defines the quality of the final product.

This fundamental change requires moving away from traditional agile approaches where developers jump into coding with minimal upfront planning. The practice of writing limited specifications is no longer adequate, as LLMs will otherwise make too many assumptions about the system you are building.

Since AI agents can save us hours of routine coding, it is crucial to invest that time in writing comprehensive specifications and thoroughly considering implementation scenarios before any code is generated. This upfront investment helps reduce bugs and prevents the costly cycle of rewriting already developed functionality.

This documentation is not only valuable input for the LLM but also makes the work of other developers using the code much easier. It helps keep the codebase consistent over time and reduces the effort required to manage technical debt