Despite the rise of coding assistants, allowing them to roam free in projects can lead to a colossal amount of cognitive debt. While robots may raise PRs in this "brave new world," reviewing hundreds of lines of subtly incorrect, AI-generated code is far from fun. For personal projects, the joy comes from the process, not just the outcome.

To avoid ceding control to the "slop machine," Ankur Sethi uses a slightly comical but effective solution: asking the assistant to generate code in the chat, then manually retyping every edit himself. By using specific instructions to prevent the LLM from directly modifying files, he maintains full control over the repository state.

While this method makes him only 2x faster instead of 10x faster, the trade-off is worth it. Manually typing the code forces a slower pace that builds a deep mental model and a spatial map of the codebase. This process helps detect hallucinations and ensures a fundamental understanding of the software, preventing a future where we no longer understand our own digital infrastructure.