ramyun retweetledi

Claude Code gets better when you give it operating rules, not vibes.
This repo turns Andrej Karpathy's notes on LLM coding failures into a reusable Claude Code guideline file.
The useful part is the bias: make the agent surface assumptions, keep diffs small, avoid speculative abstractions, and define success criteria before it starts coding.
Key Ideas:
• Think before coding: ask when ambiguity changes implementation
• Simplicity first: no abstractions until needed
• Surgical changes: touch only what traces to the request
• Goal-driven execution: tests and checks become the loop target
• Cursor support: the same guidelines can be reused outside Claude Code
Good reminder: agents don't just need more context. They need constraints that make bad coding habits harder.
Link to the GitHub repo below!

English



































