$ show multica-ai/andrej-karpathy-skills
A Claude Code and Cursor plugin that delivers behavioral guidelines for AI coding assistants, derived from Andrej Karpathy's observations about LLM coding pitfalls. The core artifact is a single CLAUDE.md file (and its Cursor rules equivalent) encoding four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. The repository packages this file as a Claude Code plugin with marketplace metadata, a Cursor project rule with alwaysApply, and a reusable skill definition. The EXAMPLES.md file provides detailed before/after code examples demonstrating each principle. The entire functional surface is documentation-as-configuration — there is no application logic, data processing, or runtime code. The plugin packaging is clean with proper marketplace.json and plugin.json manifests. The Cursor rule is correctly structured with frontmatter. The examples are concrete and well-annotated, showing realistic LLM mistakes and corrections. The Chinese README translation is a nice touch for accessibility. This is essentially a configuration and documentation project that delivers genuine value through carefully crafted AI assistant instructions.
The LLM identified 4 files worth highlighting for reuse:
A dense, well-structured behavioral contract for AI coding assistants. The four principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) are each actionable with concrete tests ('Would a senior engineer say this is overcomplicated?', 'Every changed line should trace directly to the user's request'). The document is self-contained, requires no internal dependencies, and can be dropped into any project. The pattern of transforming vague instructions into verifiable goals with step-by-step verification loops is a reusable technique for prompt engineering beyond this specific repo.
Comprehensive before/after examples for each of the four principles, showing realistic LLM overengineering (Strategy Pattern for a single discount function, 300-line preference manager for a simple save) and the corrected minimal versions. Each example includes annotated problems and a clear explanation of why the complex version is wrong despite following 'best practices'. The insight that the problem is timing — adding complexity before it's needed — is a reusable mental model for code review. The examples are self-contained and could be extracted as a standalone AI coding anti-patterns reference.
Clean, well-formed Claude Code plugin marketplace manifest with proper metadata, versioning, categorization, and keyword tags. Demonstrates correct plugin packaging structure that other developers can follow. Minimal but complete — no bloat, no missing fields.
Properly structured Cursor project rule with YAML frontmatter (description, alwaysApply: true) wrapping the same behavioral guidelines. Shows how to make AI assistant instructions tool-agnostic by packaging the same content for multiple AI coding tools. The frontmatter is correctly formatted and the content is kept in sync with CLAUDE.md as documented in CURSOR.md.