Awesome Skills for LLMs

A curated collection of resources, papers, tools, and frameworks for building, composing, and deploying skills for large language models — centered on Skills ecosystem and radiating outward to the broader LLM agent capabilities landscape.

Agent Skills was introduced as composable, portable folders of instructions, scripts, and resources that loads dynamically — turning a general-purpose assistant into a specialized agent. This list tracks everything in the skills-for-LLMs ecosystem: the official Anthropic skill system, the closely related Model Context Protocol (MCP), academic research on skill acquisition and tool use, open-source agent frameworks, computer-use and GUI agents, benchmarks, and practical tutorials.

---

Contents

• Awesome Skills for LLMs
  • Contents
  • Anthropic Skills — Core Ecosystem
    • Official Announcements & Blog Posts
    • Documentation & Guides
    • GitHub Repositories
    • Courses & Webinars
  • Model Context Protocol (MCP)
    • Specification & Announcements
    • SDKs & Official Repos
    • Community MCP Resources
  • Claude Tool Use & Computer Use
    • Tool Use Resources
    • Computer Use Resources
  • Academic Papers
    • Skill Learning & Composition
    • Tool Use & Function Calling
    • Computer Use & GUI Agents
    • Web Agents & Browser Automation
    • GUI Grounding & Visual Understanding
    • OS Agents
    • Multi-Agent Collaboration
    • Surveys & Overviews
  • Open-Source Projects & Frameworks
    • Agent Frameworks
    • Browser Automation & Computer Use Agents
    • Coding Agents
  • Benchmarks & Evaluation
    • Agent Benchmarks
    • Tool Use & Function Calling Benchmarks
    • Computer Use & GUI Benchmarks
    • Web Browsing Benchmarks
    • Leaderboards & Aggregators
  • Tutorials & Educational Resources
    • Anthropic Official Tutorials
    • Community Skills Guides
    • MCP Tutorials
    • Agent Architecture Guides
    • Tool Use & Function Calling Tutorials
    • Courses
  • Related Awesome Lists
  • Key Timeline
  • Contributing
  • Citation

---

Anthropic Skills — Core Ecosystem

Official Announcements & Blog Posts

• Introducing Agent Skills — Official product launch. Skills are folders of instructions, scripts, and resources that Claude loads dynamically. Available for Pro, Max, Team, and Enterprise users. Updated Dec 18, 2025 with organization-wide management and open standard announcement. (Oct 2025, updated Dec 2025)

• Equipping Agents for the Real World with Agent Skills — Engineering deep-dive on the Agent Skills architecture: progressive disclosure, SKILL.md format, bundled code execution, and best practices. By Barry Zhang, Keith Lazuka, Mahesh Murag. (Oct 2025, updated Dec 2025)

• Skills for Organizations, Partners, the Ecosystem — Announcement of org-wide skill management, partner-built skills directory, and Skills as an open standard for cross-platform portability. (Dec 18, 2025)

• How to Create Skills: Key Steps, Limitations, and Examples — Practical guide to building Skills: defining name/description, structuring SKILL.md, writing instructions, testing, and governance best practices for teams. (Nov 2025)

• How AI Impacts Skill Formation — RCT finding that AI assistance led to 17% lower mastery scores, exploring the tension between productivity and skill development. (Jan 2026)

Documentation & Guides

• The Complete Guide to Building Skills for Claude (PDF) — Comprehensive PDF covering SKILL.md authoring, frontmatter metadata, progressive disclosure patterns, organization-level deployment, API usage, and MCP integration with partner examples (Sentry, Box, Notion, Canva). (2025)

• 2026 Agentic Coding Trends Report (PDF) — Report on how agentic AI reshapes SDLC, developer roles, and security, covering Claude Code, Skills, and the MCP ecosystem. (Early 2026)

• Build with Claude — Learning Hub — Hub page linking to Skills docs, best practices, API usage, Claude Code usage, and MCP integration.

GitHub Repositories

• anthropics/skills — Public repository for Agent Skills: official skill definitions, examples, and the marketplace. ⭐ 62k+

• anthropics/claude-code — Agentic coding tool for terminal. Supports Skills, MCP, subagents, slash commands, and hooks. ⭐ 42k+

• anthropics/claude-cookbooks — Recipes and notebooks including agent patterns, tool use examples, and reference implementations. ⭐ 28k+

• anthropics/courses — Anthropic's educational courses including tool use tutorials. ⭐ 18k+

• anthropics/prompt-eng-interactive-tutorial — Interactive prompt engineering tutorial. ⭐ 26k+

• anthropics/claude-quickstarts — Quickstart projects including the computer-use-demo Docker container. ⭐ 10k+

• anthropics/claude-agent-sdk-typescript — TypeScript Claude Agent SDK for building custom agents with MCP integration.

• anthropics/anthropic-sdk-python — Official Python SDK for the Anthropic API.

Courses & Webinars

• Agent Skills Webinar: Transform Claude from Assistant to Specialized Agent — Technical webinar by Marius Buleandra (Anthropic Applied AI) covering Skills architecture, live demos, and best practices. (Nov 2025)

• Agent Skills with Anthropic — DeepLearning.AI — DeepLearning.AI course on Agent Skills covering the open standard, composability, and cross-platform usage.

• Introduction to Model Context Protocol — Anthropic Academy — Official Anthropic course on building MCP servers and clients using Python.

• Claude Code in Action — Anthropic Academy — Free official course for integrating Claude Code into development workflows.

---

Model Context Protocol (MCP)

Specification & Announcements

• Introducing the Model Context Protocol — Original MCP launch. Open standard for connecting AI to data sources. Pre-built servers for Google Drive, Slack, GitHub, Postgres, Puppeteer. Early adopters: Block, Apollo, Zed, Replit, Sourcegraph. (Nov 2024)

• Donating MCP and Establishing the Agentic AI Foundation — MCP donated to Linux Foundation's AAIF, co-founded with Block and OpenAI. 10,000+ active MCP servers, adopted by ChatGPT, Cursor, Gemini, VS Code. 97M+ monthly SDK downloads. (Dec 9, 2025)

• MCP Specification (2025-11-25) — Latest MCP spec: JSON-RPC 2.0, tools/resources/prompts, security guidelines, async operations, statelessness, server identity.

• One Year of MCP — November 2025 Spec Release — Anniversary blog detailing 2025-11-25 spec features: tasks, async operations, governance updates.

• MCP Joins the Agentic AI Foundation — Blog post about MCP's donation to Linux Foundation's AAIF.

• Code Execution with MCP — How to use code execution to interact with MCP servers more efficiently, reducing token overhead from tool definitions. (Nov 4, 2025)

SDKs & Official Repos

• modelcontextprotocol/modelcontextprotocol — Main specification and documentation repo. Schema in TypeScript, available as JSON Schema.

• modelcontextprotocol/servers — Official and community MCP server implementations: GitHub, GitLab, Google Drive, Google Maps, PostgreSQL, Puppeteer, Redis, Sentry, Slack, SQLite, and hundreds more.

• modelcontextprotocol/python-sdk — Official Python SDK for building MCP servers and clients.

• modelcontextprotocol/typescript-sdk — Official TypeScript SDK. v2 with Streamable HTTP, Express/Hono integrations.

• modelcontextprotocol/go-sdk — Official Go SDK, maintained in collaboration with Google.

• modelcontextprotocol/csharp-sdk — Official C#/.NET SDK, maintained in collaboration with Microsoft.

• modelcontextprotocol/kotlin-sdk — Official Kotlin SDK, maintained in collaboration with JetBrains.

• modelcontextprotocol/registry — Community-driven registry service for discovering MCP servers. Launched preview Sep 2025.

• modelcontextprotocol/ext-apps — MCP Apps Extension — standard for interactive UIs embedded in AI chatbots via MCP servers. Supports React, Vue, Svelte, Solid, Preact.

• modelcontextprotocol/use-mcp — Lightweight React hook for connecting to MCP servers.

Community MCP Resources

• punkpeye/awesome-mcp-servers — Comprehensive curated collection of MCP servers: production-ready and experimental, covering file access, databases, API integrations, and more. ⭐ 15k+

• wong2/awesome-mcp-servers — Curated list of MCP servers with official integrations, reference servers, and community servers by category. ⭐ 10k+

• microsoft/mcp — Catalog of official Microsoft MCP server implementations including Azure services, DevOps, M365 Agents Toolkit, Fabric, and Sentinel.

---

Claude Tool Use & Computer Use

Tool Use Resources

• Introducing Advanced Tool Use on the Claude Developer Platform — Three beta features: Tool Search Tool, Programmatic Tool Calling, and Tool Learning. 85% token reduction; Opus 4.5 improved from 79.5% to 88.1% accuracy. (Nov 24, 2025)

• Writing Effective Tools for Agents — With Agents — Best practices for designing MCP tools and agent tools: ergonomic design, namespacing, evaluation-driven improvement, using Claude to optimize its own tools. (2025)

• Building Agents with the Claude Agent SDK — Claude Agent SDK (renamed from Claude Code SDK): tools as primary building blocks, MCP integration, bash tool, code generation patterns. (Sep 29, 2025)

• Building Effective AI Agents — Foundational guide on agentic system patterns: prompt chaining, routing, parallelization, orchestrator-workers. Distinguishes workflows vs agents. (Dec 2024)

• Demystifying Evals for AI Agents — Guide to building evaluations for agents including coding agents, research agents, and computer use agents. (2025)

• Tool Use with Claude — Overview — Main documentation hub: client tools, server tools (web search, text editor, code execution, computer use), structured outputs, MCP integration.

• How to Implement Tool Use — Step-by-step implementation guide covering tool definitions, tool_use responses, tool_result handling, parallel tool calls, and token-efficient tools.

• Programmatic Tool Calling — Documentation for the advanced-tool-use-2025-11-20 beta: Claude executes tools programmatically via code execution.

Computer Use Resources

• Developing a Computer Use Model — Research insights on training Claude for computer use: generalization from simple software, safety considerations, and RSP assessment. (Aug 2025)

• Monitoring Computer Use via Hierarchical Summarization — Safety research on monitoring Computer Use API activity using hierarchical summarization to detect harmful behaviors at scale. (2025)

• Computer Use Tool Documentation — Official API docs for computer use: computer_20251124 (Opus 4.5/4.6), computer_20250124 (Sonnet 4.5), screenshot capture, mouse/keyboard control.

• API Release Notes — Changelog documenting computer_20250124 tool version (Jan 2025), bash_20250124, text_editor_20250124, token-efficient tool use.

---

Academic Papers

Skill Learning & Composition

• SAGE: Reinforcement Learning for Self-Improving Agent with Skill Library — Jiongxiao Wang et al. Proposes SAGE (Skill Augmented GRPO for self-Evolution), an RL framework that systematically incorporates a skill library into agent training. Achieves 8.9% higher task completion while requiring 26% fewer steps and 59% fewer tokens on AppWorld. (Dec 2025)

• CUA-Skill: Develop Skills for Computer Using Agent — Tianyi Chen et al. (Microsoft). Skill-centric framework encoding human computer-use knowledge as reusable, parameterized skills with execution and composition graphs. CUA-Skill Agent achieves SOTA 57.5% success rate on WindowsAgentArena. (Jan 2026)

• Agentic Proposing: Enhancing LLM Reasoning via Compositional Skill Synthesis — Zhengbo Jiao et al. Models problem synthesis as a goal-driven process with a specialized agent that dynamically selects and composes modular reasoning skills from a skill library. A 30B solver achieves 91.6% on AIME 2025. (Feb 2026)

• When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail — Xiaoxiao Li. Investigates "compiling" multi-agent systems into single-agent skill libraries, finding substantial reductions in token usage and latency while maintaining accuracy. Discovers a phase transition in skill selection accuracy at a critical library size. (Jan 2026)

• Self-Distillation Enables Continual Learning — Idan Shenfeld, Mehul Damani et al. Studies whether pretrained LLMs can acquire new, narrowly defined skills (science Q&A, tool use, medical reasoning) without degrading existing abilities, using self-distillation. (Jan 2026)

Tool Use & Function Calling

• Tool Learning with Large Language Models: A Survey — Changle Qu et al. Comprehensive survey organizing tool learning into four key stages: task planning, tool selection, tool calling, and response generation. Published in Frontiers of Computer Science, Vol. 19(8), 2025.

• ReTool: Reinforcement Learning for Strategic Tool Use in LLMs — Jiazhan Feng et al. Tool-augmented RL framework interleaving real-time code execution within natural language reasoning. Achieves 67% accuracy on AIME2024 with 400 training steps. Demonstrates emergent "aha moment" behaviors in tool use. (Apr 2025)

• FunReason: Enhancing LLMs' Function Calling via Self-Refinement Multiscale Loss — Bingguang Hao et al. Novel framework combining automated data refinement with Self-Refinement Multiscale Loss (SRML), achieving performance comparable to GPT-4o while mitigating catastrophic forgetting. (May 2025)

• Improving LLM Function Calling via Guided-Structured Templates — Hy Dang et al. Curriculum-inspired framework using structured reasoning templates for function calling, achieving 3-12% relative improvements over strong baselines. Shows free-form CoT is insufficient for structured function calling. (EMNLP 2025)

• LLMOrch: Efficient Function Orchestration for Large Language Models — Xiaoxia Liu et al. Automated, parallel function calling by modeling data relations (def-use) among function calls. Achieves 2× speedup over SOTA techniques. (Apr 2025)

• Function Calling in LLMs: Industrial Practices, Challenges, and Future Directions — Comprehensive ACM survey covering industrial practices in LLM function calling, including training approaches, evaluation frameworks, and future directions. ACM Computing Surveys, 2025.

Computer Use & GUI Agents

• UI-TARS: Pioneering Automated GUI Interaction with Native Agents — Yujia Qin et al. (ByteDance/Tsinghua). End-to-end native GUI agent achieving SOTA on 10+ GUI benchmarks (OSWorld 24.6, AndroidWorld 46.6) through enhanced perception, unified action modeling, and System-2 reasoning. (Jan 2025)

• UI-TARS-2: Advancing GUI Agent with Multi-Turn Reinforcement Learning — Haoming Wang et al. (ByteDance). Data flywheel for scalable generation, stabilized multi-turn RL, hybrid GUI+terminal environment. Achieves 47.5 on OSWorld, 88.2 on Online-Mind2Web, 73.3 on AndroidWorld. (Sep 2025)

• Agent S2: A Compositional Generalist-Specialist Framework for Computer Use — Saaket Agashe et al. (UC Santa Cruz). Compositional framework with Mixture-of-Grounding for precise GUI localization. Achieves 18.9% and 32.7% relative improvements over Claude Computer Use and UI-TARS on OSWorld. (Apr 2025)

• OpenCUA: Open Foundations for Computer-Use Agents — Xinyuan Wang et al. (XLANG Lab/HKU). Comprehensive open-source framework with AgentNet (first large-scale CUA dataset spanning 3 OSes, 200+ apps). OpenCUA-72B achieves 45.0% on OSWorld-Verified (SOTA among open-source). (NeurIPS 2025 Spotlight)

• ShowUI: One Vision-Language-Action Model for GUI Visual Agent — Kevin Qinghong Lin et al. Lightweight 2B vision-language-action model achieving 75.1% accuracy in zero-shot screenshot grounding with only 256K training data. (CVPR 2025)

• ShowUI-Aloha: Human-Taught GUI Agent — Yichun Zhang et al. Human-taught desktop agent via a record-parse-learn paradigm. Transforms raw interactions into semantically grounded teaching trajectories. (Jan 2026)

• AFRAgent: Adaptive Feature Renormalization Based GUI Agent — Neeraj Anand et al. Less than one-fourth the size of nearest competitor while achieving SOTA on Meta-GUI and AITW benchmarks. (Nov 2025)

Web Agents & Browser Automation

• The BrowserGym Ecosystem for Web Agent Research — Thibault Le Sellier De Chezelles et al. (ServiceNow/CMU/McGill). Unified gym-like environment for standardized web agent evaluation. First large-scale, multi-benchmark experiment comparing 6 LLMs across 6 benchmarks. (TMLR Feb 2025)

• BrowserAgent: Web Agents with Human-Inspired Browsing Actions — ReAct-style reasoning framework with explicit memory for multi-turn web interactions. With only 5.3K training samples, outperforms Search-R1 on Open-QA tasks. (Oct 2025)

• Evaluating Long-Context Reasoning in LLM-Based WebAgents — Andy Chung et al. Benchmark for long-context reasoning in WebAgents. Finds dramatic performance degradation (40-50% → <10%) as context length increases. (Dec 2025)

• Building Browser Agents: Architecture, Security, and Practical Solutions — Examines production-grade browser agent architecture, covering context management, safety boundaries, and the choice between generalization and specialization. (Nov 2025)

GUI Grounding & Visual Understanding

• UGround: Universal Visual Grounding for GUI Agents — Boyu Gou et al. (OSU NLP). Trains on 10M GUI elements from 1.3M screenshots — the largest GUI visual grounding dataset. Outperforms existing models by up to 20% absolute. (ICLR 2025 Oral)

• Enhancing Visual Grounding for GUI Agents via Self-Evolutionary RL — Xinbin Yuan et al. RL-based framework where a 7B model achieves 47.3% on ScreenSpot-Pro, outperforming UI-TARS-72B by 24.2% with only 3K training samples. (May 2025)

• Scaling Computer-Use Grounding via UI Decomposition and Synthesis (Jedi) — Tianbao Xie et al. (XLANG Lab). OSWorld-G benchmark (564 samples) and Jedi dataset (4M grounding examples). Improves OSWorld agentic success from 5% to 27%. (NeurIPS 2025 Spotlight)

• ScreenSpot-Pro: GUI Grounding for Professional High-Resolution Computer Use — Kaixin Li et al. Benchmark spanning 23 applications across 5 industries and 3 OSes. Best models initially achieved only 18.9%. (Apr 2025)

• [UI-R1: Enhancing GUI Action Prediction by Reinforcement Learn