Sunday's GitHub trending tells a coherent story: the developer community is done treating AI coding agents as magic boxes and is now actively engineering how those agents think, what they know, and what they can do. The top repos today span behavior constraints, structured knowledge, plugin ecosystems, and foundational education — infrastructure for a more disciplined era of AI-assisted development.
multica-ai/andrej-karpathy-skills: Four Rules That Tame Claude's Bad Habits
The day's biggest mover earned over 3,500 stars for a single CLAUDE.md file. Drawing on observations Andrej Karpathy has shared about common LLM coding failure modes, the file codifies four principles: reason explicitly before acting, prefer the simplest solution that works, make surgical edits rather than wholesale rewrites, and drive toward goals rather than executing steps blindly. It installs as a Claude Code plugin in seconds and immediately changes how the agent approaches ambiguous tasks — less assumption-making, more deliberate thought. The velocity suggests developers have been waiting for exactly this kind of opinionated behavioral layer.
github.com/multica-ai/andrej-karpathy-skills
anthropics/claude-plugins-official: A Curated Marketplace for Claude Code
Anthropic shipped an official, Anthropic-maintained plugin directory for Claude Code, pulling in over 2,100 stars on its first day. The repo establishes a standardized plugin format — plugin.json, optional MCP server integration, slash commands, agents, and skill packs — and provides a single place to discover both Anthropic-built and community-submitted extensions. External plugins go through a quality and security review before listing, which addresses the trust problem that informal sharing couldn't solve. Installation is a one-liner through Claude Code's plugin system, and the reference implementation gives third-party developers a clear target to build against.
github.com/anthropics/claude-plugins-official
colbymchenry/codegraph: Code Knowledge Graphs That Run Locally
Still near the top after multiple days, codegraph gained another 2,400+ stars as more developers discovered its approach to giving AI agents structured understanding of a codebase. Rather than relying on vector search over raw files, it pre-indexes code into a knowledge graph that tracks relationships between modules, functions, and dependencies — letting an agent ask relational questions, not just keyword ones. The emphasis on local operation without a cloud dependency is a deliberate design choice: the graph is built once and queried offline, keeping sensitive code out of external services. Its continued rise alongside Understand-Anything (yesterday's digest) confirms that structured code comprehension is the category to watch.
github.com/colbymchenry/codegraph
rohitg00/ai-engineering-from-scratch: A Curriculum That Builds the Full Stack
With 1,500 stars today, this open-source curriculum takes a notably different approach from most AI learning resources: it covers 435 lessons across 20 phases, starting from mathematical foundations and working up through autonomous systems and AI ethics. Crucially, the course builds algorithms from first principles before introducing frameworks, so learners understand what tools are actually doing rather than just calling APIs. Each lesson produces a reusable artifact — a prompt, skill, agent, or MCP server — meaning the curriculum doubles as a growing personal toolkit. For engineers who feel like they skipped steps on the way to shipping AI features, this is a serious catch-up path.