← Back to digest
HN Threads

Retype your LLM code by hand?

What's this about?

Ankur Sethi's practice: never paste AI-generated code — retype it line by line, so every line passes through your understanding. The claim: you keep the speed of generation while avoiding "cognitive debt", the codebase-you-don't-understand problem.

What exactly is "cognitive debt"?

By analogy with technical debt: every pasted-but-not-understood block is borrowed comprehension. It compounds — one day you must debug code nobody at your company has actually read. Retyping is a forcing function for reading.

What's the community saying?

HN 398 points, 337 comments — unusually split.
  1. Mockery: "a new form of prayer for those who can't break their addiction to code generators."
  2. Pragmatists: this throws away most of the speed advantage — review rigorously instead of performing typing rituals.
  3. Veterans: it's the classic "type out the textbook examples" study technique, which genuinely works for learning.
  4. A middle camp retypes only unfamiliar-domain code and pastes boilerplate — comprehension where it matters.
← Nightcrawler: AI pentesting agent on a smartphoneAirLLM: 70B inference on a 4GB GPU →