Prevent cognitive debt by manually retyping LLM-generated code

Developer Ankur Sethi has described a personal workflow that lets him keep using AI coding assistants on his own projects without handing over control of his code. In an essay on his blog he says he still uses the tools on personal projects, but that one-shotting entire features leaves him unsatisfied and disoriented — what he enjoys is fast-forwarding through the boring parts.

The problem: cognitive debt

Letting an assistant roam free, he writes, leaves him with “a colossal amount of cognitive debt”. Reading the Django documentation to figure out how to add tagging to his website may be boring, but he still wants to understand how it works: “Just because a problem is boring doesn’t mean I want to fully offload my understanding of the solution to a machine.” Reviewing every line, the approach most developers are expected to take in 2026, is not a fix either: “Robots raise PRs, humans review them. It’s a brave new world.” He says poring over hundreds of lines of “overly-defensive, badly-commented, subtly incorrect code” is not fun — acceptable sometimes for an employer, but never for personal projects, which must be fun above all else.

A deliberately inefficient workflow

His solution is, in his own words, “grossly inefficient and perhaps slightly comical”: he asks his coding assistant to generate code in the chat, then makes all the edits himself. The instructions sit in the agent files in all his personal projects. He estimates the approach makes him faster than working without LLMs at all, but slower than developers willing to let the machine think for them — “instead of being 10x faster, I’m probably only 2x faster”.

What he gains is understanding. Manually typing each line builds a mental model of how the code fits into his existing codebase; if he doesn’t understand an API or algorithm he can stop and look it up. Typing slows him down, so he is more likely to catch hallucinations or bad design choices, and he can reorganize and comment the code as he goes. Most importantly, it builds a “spatial map” of the codebase — he knows where every bit of functionality lives, which also makes his future prompts to the LLM better.

Why it matters

Sethi compares the practice to the old advice given to programmers learning to code: never copy and paste code — retype examples from books and blog posts so you understand them. “It might not be the most efficient way to work with an LLM, but I value comprehension over productivity,” he writes, adding that he has been working this way for a few months and plans to continue. He worries that the software industry is taking on a large amount of cognitive debt that will have to be paid back soon: “There will come a time when we no longer understand how large parts of our digital infrastructure are put together.”