A developer has adopted an unconventional workflow for using coding assistants on personal projects: requesting AI-generated code in chat, then manually retyping every line into their editor rather than copying it directly.
According to the author, this approach stems from dissatisfaction with two common alternatives. Allowing an LLM to generate entire features leaves them “unsatisfied and disoriented,” while reviewing hundreds of lines of AI-generated code—the expected practice in what they call “this cursed year of 2026”—feels tedious and unenjoyable on personal projects where the process matters more than the outcome.
The manual retyping strategy trades raw speed for comprehension. The author estimates it makes them roughly 2x faster than working without LLM assistance, rather than the often-cited 10x productivity gains. However, the slowdown produces measurable cognitive benefits. As they type each line, they build a mental model of how the code functions and integrates with their existing codebase. If they encounter unfamiliar APIs or algorithms, they can pause to research or ask the LLM for clarification.
The physical act of typing also serves as a quality filter. The author reports being more likely to spot hallucinations or poor design choices while manually entering code, allowing them to refactor, reorganize, and add comments as they go. Critically, the process creates what they describe as a “spatial map” of their codebase—they know exactly where each component lives and can quickly navigate to make changes later.
The author draws a parallel to traditional coding education. Experienced programmers historically advised learners to type out examples from books and adapt code from blog posts rather than copying and pasting, ensuring genuine understanding. They argue that manually entering LLM-generated code replicates this same learning process.
The author expresses concern about broader industry trends, suggesting that widespread reliance on AI-generated code without comprehension could create significant “cognitive debt” in digital infrastructure. They frame their personal practice as a safeguard: understanding the software they ship represents a professional responsibility. They have been using this workflow for several months and plan to continue it.
Key facts
- The developer manually retypes LLM-generated code rather than copying it directly
- This approach is estimated to provide roughly 2x speed improvement over traditional coding, significantly less than claimed 10x gains
- Manual typing allows the developer to spot hallucinations, understand APIs, refactor code, and build a mental map of their codebase
- The author compares the practice to traditional advice for new programmers to type out code examples for better learning