Prevent cognitive debt by manually retyping LLM-generated code

A developer expresses dissatisfaction with using coding assistants to generate entire features for personal projects. The author finds that offloading too much to a machine leaves them with an overwhelming amount of cognitive debt. They fear that understanding the inner workings of their codebase becomes too difficult. Despite the efficiency gains from using LLMs, the author prefers a slower approach that allows for a deeper comprehension of their code.

The developer's solution is to manually edit every line of code generated by the assistant. This manual process forces them to engage with the code, build a mental model of how it works, and understand the APIs or algorithms used. It also helps detect errors and hallucinations in the AI-generated code, allowing them to clean it up and adapt it to their own coding style.

By manually typing the code, the author develops a spatial map of their codebase, making it easier to navigate and make changes in the future. This workflow provides a balance between the speed benefits of AI assistance and the deep understanding that comes from manual coding. The author has been using this method for several months and plans to continue it, fearing that the software industry is accumulating significant cognitive debt that could be difficult to reverse.

While this method may not be the most efficient, the author values comprehension over productivity, prioritizing their understanding of the software they create.