Verified:GitHub

karpathy

Enforce minimal code changes with explicit assumptions and verifiable success criteria.

56|20|Updated Mar 26, 2026

Search Tags:#coding #assumptions #guidelines #simplicity #deterministic execution #surgical edits #code change verification

Ready to run

npx skills add https://github.com/markster-public/markster-os --skill karpathy-markster-public

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: karpathy
Source: https://github.com/markster-public/markster-os/tree/main/skills/karpathy
Command: npx skills add https://github.com/markster-public/markster-os --skill karpathy-markster-public

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

It prevents LLM coding mistakes by enforcing strict simplicity, explicit assumptions, and minimal, goal-driven changes that reduce unintended side effects.

Core Features & Use Cases

Think Before Coding: Forces the model to surface uncertainty and ask when requirements are ambiguous instead of guessing.
Simplicity First: Limits output to the smallest correct solution to avoid overengineering.
Surgical Changes: Restricts edits to only what the user asked for, and cleans up any new orphans introduced by the change.

Use case: You’re doing a complex feature update and want predictable, low-risk diffs—especially when the existing codebase should not be refactored or “improved” beyond the request.

Quick Start

Enable the karpathy skill and ask your AI to perform a focused code change with explicit assumptions and defined success criteria.

Frequently Asked Questions about karpathy

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I stop AI from overengineering and refactoring code I didn't ask to change?▼

To prevent hidden assumptions in AI code generation, use a discipline that forces the model to surface uncertainty and ask questions when requirements are ambiguous. This ensures explicit assumptions are stated before any code is written, reducing unpredictable behavior.

What is the best way to get minimal, predictable diffs for complex feature updates?▼

The best way to get minimal, predictable diffs is to enforce surgical code edits with defined success criteria tied to measurable verification. This approach limits changes to the specific request and cleans up any new orphans introduced by the modification.

Why does my AI assistant guess requirements instead of asking for clarification?▼

To verify code change success and ensure deterministic execution, tie your success criteria to measurable verification outcomes. This enforces strict coding discipline, requiring explicit assumptions and minimal code that satisfies the request predictably.

Can I apply surgical code edits to an existing codebase without triggering unwanted refactors?▼

Yes, you can apply surgical code edits to an existing codebase without unwanted refactors by enforcing strict simplicity and minimal, goal-driven changes. This discipline restricts edits to your request and cleans up orphaned code without improving beyond scope.