Ankur Sethi, a developer writing on August 2, 2026, said he still uses coding assistants for personal projects but avoids letting them directly edit files because the workflow leaves him with "cognitive debt," or code he owns without fully understanding. The post, which drew 247 points and 198 comments on Hacker News, argues that LLMs can speed up boring programming tasks while still weakening a developer's mental grasp of a codebase if they generate whole features or large pull requests.

Sethi said his preferred method is deliberately inefficient: he asks the coding assistant to propose code in chat, then manually types every edit into his editor. His standing instructions tell agents never to create, edit, move, rename, or delete project files unless explicitly asked, and never to run commands that modify files, install dependencies, or change repository state without explicit permission. The assistant should show proposed edits and commands in chat, while explanations of syntax, APIs, programming concepts, or implementation details are withheld unless requested.

Sethi estimates the workflow makes him about 2x faster than working without LLMs, not 10x faster than people who let the machine think for them. He says manually typing every line helps him build a mental model of how the code works, notice hallucinations or bad design choices, and clean up, reorganize, refactor, or comment as he goes. He also says the practice builds a spatial map of the codebase, making it easier to know where functionality lives and to prompt the LLM more precisely later.

Sethi compares the approach to old advice for learners: never copy and paste code without typing, running, and adapting it. He says he has used this workflow for a few months and plans to continue because he values comprehension over productivity. He warns that the software industry may be accumulating cognitive debt in digital infrastructure, while his personal goal is to fully understand the software he releases.

Ankur Sethi, a developer writing on August 2, 2026, said he still uses coding assistants for personal projects but avoids letting them directly edit files because the workflow leaves him with "cognitive debt," or code he owns without fully understanding. The post, which drew 247 points and 198 comments on Hacker News, argues that LLMs can speed up boring programming tasks while still weakening a developer's mental grasp of a codebase if they generate whole features or large pull requests.

Sethi said his preferred method is deliberately inefficient: he asks the coding assistant to propose code in chat, then manually types every edit into his editor. His standing instructions tell agents never to create, edit, move, rename, or delete project files unless explicitly asked, and never to run commands that modify files, install dependencies, or change repository state without explicit permission. The assistant should show proposed edits and commands in chat, while explanations of syntax, APIs, programming concepts, or implementation details are withheld unless requested.

Sethi estimates the workflow makes him about 2x faster than working without LLMs, not 10x faster than people who let the machine think for them. He says manually typing every line helps him build a mental model of how the code works, notice hallucinations or bad design choices, and clean up, reorganize, refactor, or comment as he goes. He also says the practice builds a spatial map of the codebase, making it easier to know where functionality lives and to prompt the LLM more precisely later.

Sethi compares the approach to old advice for learners: never copy and paste code without typing, running, and adapting it. He says he has used this workflow for a few months and plans to continue because he values comprehension over productivity. He warns that the software industry may be accumulating cognitive debt in digital infrastructure, while his personal goal is to fully understand the software he releases.