Skill Discovery May 27, 2026: 140K Stars, Code Graphs, and Microsoft Goes Official

Four Skills, Four Different Theories

Today's discovery is compact but high-signal. Each of the four skills represents a different theory about what's missing from AI-assisted development.

One bets on behavior rules. One bets on structural knowledge. One bets on domain expertise. One bets on interoperability. All four have traction — and they don't overlap.

---

1. Karpathy Guidelines (multica-ai) — 140,000 Stars

The most-starred skill in today's list isn't a framework or a toolchain. It's four rules in a markdown file.

This skill distills Andrej Karpathy's observations about LLM coding behavior into explicit, enforceable directives:

1. Think before coding — write a short plan before touching any file
2. Prefer simplicity — resist the urge to abstract, generalize, or add layers
3. Make targeted changes — touch only what needs to change; don't refactor opportunistically
4. Verify by outcome — don't declare success until the result works, not just compiles

140,000 stars in a short window is one of the sharpest signals the ecosystem has produced. It means developers recognized these patterns immediately — because they'd already seen their AI do exactly these things.

The multica-ai version packages the original CLAUDE.md into a proper installable skill with .cursor/rules and plugin.json support. One install, all four tools.

Install: npx skills add multica-ai/andrej-karpathy-skills

→ Full details & install guide

---

2. CodeGraph — 27,900 Stars

There's a specific failure mode that every developer using AI on a real codebase has hit: the agent starts grep-ing, reading files one by one, building context through repeated tool calls — burning tokens and time before it's even understood the problem.

CodeGraph takes a different approach. It parses your codebase once and builds a structured knowledge graph: functions, classes, imports, call relationships, file dependencies. When an agent needs to understand how something works, it queries the graph instead of exploring the filesystem.

The practical results are meaningful. Less token usage per task. Fewer tool calls. Faster time-to-answer on "how does X work" questions. The 27.9K stars suggest this addresses a real pain point, not a theoretical one.

Works 100% locally across 19+ languages. Integrates as a CLI tool, an MCP server, or directly as an Agent Skill.

Install: npx skills add colbymchenry/codegraph

→ Full details & install guide

---

3. .NET Agent Skills — 2,575 Stars

Microsoft's .NET team has published an official skill library — and it's the most technically thorough domain-specific skill collection we've seen from a major platform team.

12 plugins, 80+ granular skills, covering the specific scenarios where general-purpose AI tools consistently hallucinate or fall back to outdated patterns:

• Core development — modern .NET idioms, not Java-in-C# habits
• Entity Framework — correct data access patterns, migrations, query optimization
• dotnet-trace / diagnostics — performance profiling workflows
• MSBuild — build system configuration that actually reflects how .NET projects work
• MAUI — cross-platform mobile development specifics
• Semantic Kernel — Microsoft's own AI integration library, which AI tools obviously can't know well without being taught

The verified badge reflects that this comes directly from the people who build .NET — not a community approximation of their practices.

Install: npx skills add dotnet/skills

→ Full details & install guide

---

4. NotebookLM Python API & Skill — 14,600 Stars

Google NotebookLM is one of the most capable research tools available, but it's locked behind a web UI that doesn't expose an API. This skill changes that.

The Python library gives programmatic access to NotebookLM's core capabilities — notebook creation, source document management, audio overview generation, research queries — including features the official interface doesn't surface. All of it callable from Python scripts, CLI pipelines, or directly from Claude Code and Codex agents.

The practical use case: AI research workflows that bridge NotebookLM's synthesis capabilities with agent-driven automation. Ask Claude to create a notebook from a set of papers, extract key arguments, and generate a structured summary — without touching a browser.

14.6K stars for a tool that talks to a platform that officially has no API is a strong signal that this fills a real gap.

Install: npx skills add teng-lin/notebooklm-py

→ Full details & install guide

---

What These Four Have in Common

They all solve the same underlying problem from different angles: AI agents don't have the right context by default.

Karpathy Guidelines fix behavioral context — the rules AI should follow but doesn't. CodeGraph fixes structural context — knowledge of your codebase that AI shouldn't have to rebuild each time. .NET Skills fix domain context — expertise in a specific technology stack. NotebookLM fixes research context — access to synthesized knowledge that's otherwise locked in a closed platform.

The skills that are winning aren't the ones that teach AI to do new things. They're the ones that give AI the context it needs to do familiar things correctly.

---

All 4 Skills

• Karpathy Guidelines (multica-ai) — 140,000 ⭐ — Behavior rules for coding agents
• CodeGraph — 27,900 ⭐ — Pre-indexed code knowledge graph
• .NET Agent Skills — 2,575 ⭐ — Microsoft official .NET skill library
• NotebookLM Python API — 14,600 ⭐ — Programmatic access to Google NotebookLM

Or browse the full catalog →

---

Skill Discovery runs daily. Subscribe to get the next edition.