To Avoid 'Cognitive Debt,' This Developer Retypes Every Line of AI Code by Hand

Blogger Ankur Sethi describes a deliberately inefficient workflow for coding assistants: rather than letting an LLM edit his projects directly or reviewing the pull requests it raises, he has the assistant produce code only in the chat window, then retypes every line into his editor by hand. His agent instruction files enforce this rule. The point isn’t speed—he estimates it makes him roughly 2x faster instead of the 10x some claim—but comprehension. Manually transcribing the code forces him to slow down enough to catch hallucinations and questionable design choices, refactor as he goes, and build a spatial map of where each piece of functionality lives.

Sethi frames this against what he sees as the default mode of 2026 development, where machines open PRs and humans grudgingly review them. He finds reviewing AI output tedious and unrewarding, especially for personal projects that are supposed to be fun, and argues that offloading understanding to the “slop machine” is a poor trade. He likens hand-typing generated code to the old advice given to beginners: never copy-paste from a book or forum, type it out and adapt it so you actually understand it.

The broader worry driving the post is industry-scale cognitive debt—a near future in which large parts of digital infrastructure are assembled by tools that nobody on the team fully understands. Sethi concedes he can’t change the trajectory of the whole field, but insists on understanding the software he personally ships, calling anything less professional malpractice.