Prevent cognitive debt by manually retyping LLM-generated code

TL;DR Highlight

Blocking agent file edits and retyping code by hand improved one dev's understanding, at about 2x speed.

Who Should Read

A solo developer uneasy because code from coding assistants has left them not knowing what lives where in their own project

Core Mechanics

Make the agent show edits only in chat
- The agents file forbids creating, editing, moving, renaming or deleting files unless asked.
- Proposed edits appear in chat and the author types them in.
Run state-changing commands yourself
- Commands that install dependencies or change repo state are shown in chat instead.
- The author runs them manually.
About 2x faster, not 10x
- The author says this is faster than using no LLM.
- It is slower than letting the machine do everything.
Typing surfaces hallucinations and bad design
- Slowing down makes errors more likely to be noticed.
- The author also refactors and comments while typing.
You get a spatial map of the codebase
- The author knows where each piece of functionality lives.

Community Reactions

Where are the efficiency gains?
- One commenter asked why not drop the LLM if you think, read, retype and fix anyway.
- Another said retyping to understand makes you a code monkey.
Only self-written code builds the skill
- One commenter said retyping is a no, but typing your own code is a yes.
- They described learning Ruby and Go by retyping others' example solutions.
Experience retyping code
- One commenter retyped an inherited codebase and fixed many bugs while becoming an expert quickly.
- Another used this approach, then stopped and blogged about it.
Rethink the understanding level
- One commenter reviews AI code like a teammate's and accepts, edits or requests revisions.
- Another suggested managing LLMs as agents, like a manager who doesn't know every line.

How to Apply

Add the no-edit rules to your agents file
- Today, paste the author's three instructions into AGENTS.md or CLAUDE.md in a personal project.
- Ask for one small feature and type the code from chat yourself.
Ask about unfamiliar APIs while typing
- Stop when an API or algorithm is unclear and ask the LLM to explain it.
- Rename and restructure to your taste as you type.
Scope where you use this
- Use it for personal projects or core modules you must understand.
- Keep your normal review flow for simple parts and compare the speed cost.

Try It

copy & run

# Paste into AGENTS.md / CLAUDE.md for Claude Code etc. (the author's instructions)

I want to understand every line of code that goes into this project. Never create, edit, move, rename, or delete project files unless I explicitly ask you to do so. Instead, show me every proposed edit in the chat so I can type it in manually.

Do not run commands that modify project files, install dependencies, or change repository state unless I explicitly request that action. Instead, show me those commands in the chat so I can run them manually.

I'm an experienced developer. Do not explain syntax, APIs, programming concepts, or implementation details unless explicitly asked.

# Next: ask "Show me the code to add {{feature}}" and type it in yourself

Terminology

cognitive debtDebt from keeping AI-written code you don't understand. You pay it later when you modify or debug it.

agents fileA config document giving a coding agent per-project rules, like a team rulebook for a new hire.

one-shotHaving the model build an entire feature from a single instruction.