In April 2026 one GitHub repo hit 179,000 stars — not a framework or a model, but a single AGENTS.md file with four behavioral rules derived from Andrej Karpathy’s observations about LLM coding pitfalls. The repo: multica-ai/andrej-karpathy-skills.
Karpathy is a founding member of OpenAI and former head of AI at Tesla, widely recognized for his work on deep learning, computer vision, and LLMs. He’s also one of the most followed voices in AI on practical agentic coding workflows.
His diagnosis zeroes in on three failure modes — and the fix is four principles that bias the agent toward caution over speed:
| # | Failure Mode | Description | Fix |
|---|---|---|---|
| 1 | Silent Assumptions | ”They make wrong assumptions on your behalf and just run along with them without checking” | → Line 1: Think Before Coding |
| 2 | Overcomplication | ”They implement a bloated construction over 1,000 lines when 100 would do” | → Line 2: Simplicity First |
| 3 | Side-Effect Edits | ”They change/remove comments and code they don’t sufficiently understand as side effects” | → Line 3: Surgical Changes |
Line 4 goes beyond fixing — it unlocks a new capability for agents to work independently.
Here’s the full mapping:
| Principle | Problem It Solves | The One-Liner |
|---|---|---|
| Think Before Coding | Wrong assumptions, hidden confusion, missing tradeoffs | Don’t assume. Don’t hide confusion. Surface tradeoffs. |
| Simplicity First | Overcomplication, bloated abstractions | Minimum code that solves the problem. Nothing speculative. |
| Surgical Changes | Orthogonal edits, touching code you shouldn’t | Touch only what you must. Clean up only your own mess. |
| Goal-Driven Execution | Vague plans with no verification | Define success criteria. Loop until verified. |
⚠️ Tradeoff: These guidelines bias toward caution over speed. For trivial tasks (typo fixes, obvious one-liners), skip the rigor. The goal is reducing costly mistakes on non-trivial work, not slowing down simple changes.
Paste this into your AGENTS.md:
- Think before coding. State assumptions, surface tradeoffs, push back when warranted.
- Simplicity first. Minimum code that solves the problem. Nothing speculative.
- Surgical changes. Touch only what you must. Clean up only your own mess.
- Goal-driven execution. Define success criteria. Loop until verified.
Source article: The 4 Lines Every AGENTS.md Needs by Yanli Liu, published in Level Up Coding (Apr 2026).