Andrej Karpathy
Andrej Karpathy is an AI researcher and engineer whose public comments on LLM coding behavior are used in these sources as the seed for a practical agent-instruction pattern.
This page records how he appears in the ingested sources, not a full biography.
Role In The Sources
karpathy-claude-md-hit-number-one presents Karpathy as the origin of four concise rules for making Claude Code less error-prone:
- ask rather than assume
- use the simplest solution first
- avoid unrelated edits
- flag uncertainty explicitly
andrej-karpathy-skills-repo expands those ideas into a portable instruction package for Claude Code, Cursor, and skill/plugin surfaces.
Why It Matters Here
Karpathy's framing fits this wiki's broader pattern:
- epistemic-humility: name uncertainty instead of hiding it.
- first-principles-thinking: choose the simplest solution from actual constraints.
- agentic-coding-workflows: make agents useful through bounded tasks and verification.
- agent-operating-instructions: preserve these rules in durable project files.
The key insight is that coding-agent quality is partly a model problem, but also a context-system problem. Better instructions, scoped tasks, and verification loops improve the whole work system.