Agent skills

What it is: google/agents-cli is an Apache-2.0 CLI and skill bundle that turns Codex, Claude Code, Gemini CLI, and similar assistants into practical Google Cloud agent builders, covering scaffold, evaluation, deployment, and operations instead of leaving teams to wire each platform step by hand.

Why it matters: Useful for builders who want a faster path from local agent prototype to governed Google Cloud deployment, especially when they need repeatable setup instead of a one-off demo.

What it is: wanshuiyin/Auto-claude-code-research-in-sleep is an MIT-licensed research-workflow kit that packages literature review, paper analysis, idea generation, and writing loops into reusable skills for Claude Code, Codex, and other LLM agents instead of locking the process inside one hosted app.

Why it matters: Useful for researchers and advanced builders who want portable, documented research workflows with multi-model review loops instead of one-off chat sessions.

What it is: getcrew44/crew44 is an MIT-licensed local-first desktop orchestrator that turns Claude Code, Codex, Cursor Agent, Gemini CLI, Hermes, and other coding-agent CLIs into a coordinated specialist team with per-project memory, shared SKILL.md capabilities, isolated git worktrees, explicit handoffs, and an independent goal verifier.

Why it matters: Useful for developers and small teams who already run multiple coding agents and want persistent specialist roles, parallel handoffs, and worktree isolation instead of re-explaining repo context to one generalist session on every task.

What it is: Graphify-Labs/graphify is a 79534-star, MIT Claude Code skill that turns any folder of code, SQL schemas, R scripts, shell scripts, docs, papers, images, or videos into a queryable, persistent, multimodal knowledge graph with 71.5x fewer tokens per query than re-reading the raw files. The headline value proposition is the durable multimodal + persistent + agent-friendly knowledge-graph surface:

Why it matters: Useful for AI agent developers, AI coding agent users, Claude Code users, knowledge-management practitioners, Obsidian users, documentation engineers, AI-curious readers tracking the Claude Code skill ecosystem, engineering teams wiring knowledge graphs to their codebases, and any developer wiring a multimodal, persistent knowledge graph to Claude Code -- and who can pair Graphify-Labs/graphify

What it is: Addy Osmani's agent-skills is an MIT-licensed repository of production-grade engineering skills and lifecycle commands that push coding agents toward spec-first planning, testing, review, and shipping workflows instead of raw prompt improvisation.

Why it matters: Useful for developers who want coding agents to follow a more disciplined software delivery process with clearer quality gates.

What it is: Google Managed Agents in the Gemini API is a preview cloud agent runtime that lets developers define agents with AGENTS.md and SKILL.md, then spin them up with one API call to reason, browse, run code, and manage files inside isolated Linux sandboxes with resumable session state.

Why it matters: Useful for teams that want agent execution infrastructure and sandbox management handled for them instead of stitching together their own orchestration, browser automation, file lifecycle, and remote execution stack.

What it is: Huashu Design alchaincyf/huashu-design is an MIT HTML-native Agent Skill for coding assistants that takes a single Chinese or English prompt and emits a ship-ready design — a product-launch animation, a clickable App prototype, an editable PPT, or a print-grade infographic — in 3 to 30 minutes, ships three logic-consultant defaults plus a 40-style HTML-native style library so output never

Why it matters: Useful for AI engineers, designers, founders, indie hackers, content creators, growth teams, and small teams who want Claude Code / Cursor / Codex to ship a polished product-launch animation, a clickable App prototype, an editable PPT, or a print-grade infographic from a single sentence in 3-30 minutes, because alchaincyf/huashu-design is an MIT HTML-native Agent Skill for coding assistants that

What it is: Apple's coreai-models repository provides model export recipes, Python primitives, Swift runtime utilities, and agent skills for building on-device AI with Apple's Core AI framework. The repo includes recipes for converting popular open models, reusable PyTorch building blocks, and Swift integration utilities for macOS and iOS apps.

Why it matters: Useful for Apple-platform developers moving from demos to shippable on-device AI: start with the model catalog and a small export recipe, then validate latency, memory, and OS-version requirements on real hardware before product use.

What it is: Ar9av/obsidian-wiki is a 2,546-star MIT framework that turns an Obsidian vault into a local second brain for coding agents, with cross-agent skills, a pip install path, and a file-based workflow for growing linked knowledge instead of re-prompting the same context every session.

Why it matters: Useful for builders who want agent memory to live in owned markdown files instead of disappearing into chat scrollback or a hosted memory layer.

What it is: BuilderIO/skills is an MIT-licensed catalog of small composable skills for coding agents, with a CLI installer, visual planning and recap flows, GitHub Action hooks, and reusable review or validation behaviors instead of one giant agent framework.

Why it matters: Useful for teams that want sharper code-agent planning, review, and handoff behavior without committing to a whole new agent stack.

What it is: Lathe devenjarvis/lathe is an MIT single-binary LLM tutorial generator that pairs a local Go CLI with bundled coding-agent skills (Claude Code / Cursor / Codex / Gemini CLI / OpenCode / Cline / Windsurf) — invoke /lathe build a 3D Slicer in Erlang in any coding agent and Lathe generates a hands-on, multi-part technical tutorial on demand with skills tuned to make content approachable, then

Why it matters: Useful for self-learners, students, bootcamp learners, working engineers picking up a new stack, technical leads onboarding new hires, instructors who want auto-generated homework, AI-assisted learning researchers, and any operator who wants to use an LLM to teach them a topic instead of doing the thinking for them, because Lathe is an MIT single-binary LLM tutorial generator that pairs a local

What it is: MIT open-source Claude Code skill + 8 Node.js executable scripts from Jane (xiaoerzhan) that reverse-engineers any website -- from a single-file static page to a WebGL-heavy interactive demo -- without faking it from AI hallucinations

Why it matters: Useful for AI coding-agent power users, front-end engineers, designers who remix other sites, technical writers, and any developer who has ever lost an afternoon to AI-generated clone analysis code that looked plausible but ran nothing like the original -- an MIT Claude Code skill that reverse-engineers any website with a real-source-first methodology, a 6-step decision tree, and 8 executable

What it is: MIT official Lark/Feishu CLI from the larksuite GitHub team -- Node-distributed Go binary (npm install -g @larksuite/cli, npx @larksuite/cli@latest install) that exposes the full Lark/Feishu Open Platform surface (200+ commands + 26 AI Agent Skills); 18 business domains

Why it matters: Most enterprise teams using Lark or Feishu today who want an AI agent to automate the full platform surface have been either (a) writing 200+ API calls by hand against the Lark/Feishu Open Platform (no canonical CLI surface), (b) reaching for non-MCP/non-agent-native CLI tools that require custom agent integration, or (c) maintaining separate Meegle / Calendar / Docs wrappers per domain.

What it is: MIT-licensed minimal self-evolving autonomous agent framework (13,341* at verify time); just ~3K lines of seed code + 9 atomic tools + a ~100-line agent loop; crystallizes every successful task into a reusable Skill that grows a personal skill tree over time; arXiv 2604.17091

Why it matters: Most autonomous-agent frameworks today have shipped enormous codebases -- tens of thousands of lines, large dependency trees, dozens of pre-loaded skills, complex tool-bridge interfaces -- that are hard to audit and harder to extend. lsdefine/genericagent inverts that pattern: the canonical MIT open-source minimal-seed self-evolving autonomous agent framework where the entire starting point is

What it is: shadcn/improve is the MIT-licensed agent skill (ships in the Agent Skills format used by Claude Code, Cursor, Codex, and OpenCode) that turns the most capable model into a repo auditor and plan-writer while handing execution to cheaper models: the maintainer's pitch is 'use your most capable model for the part where intelligence compounds — understanding the codebase, judging what's worth doing

Why it matters: Useful for engineering teams running Claude Code / Cursor / Codex / OpenCode who want their most capable model to do the part where intelligence compounds (audit + spec) and their cheaper models to do the part where execution speed compounds: shadcn/improve is the MIT agent skill that audits any codebase and writes prioritized, self-contained implementation plans for other agents to execute; for

What it is: Hope Agent is an MIT-licensed cross-device desktop assistant with persistent memory, reusable skills, optional service mode, and handoff across devices instead of one isolated chat window.

Why it matters: Useful for people who want a self-hosted personal assistant with real continuity across desktop, browser, and service modes rather than a one-tab chatbot that forgets how they work.

What it is: alchaincyf/huashu-md-html is an MIT agent-agnostic markdown publishing pipeline that converts PDFs, Office files, images, audio, YouTube links, and web pages into clean markdown, then turns that markdown back into polished HTML, archival markdown, or publisher-grade DOCX with one reusable skill pack.

Why it matters: Useful for teams that need AI-generated or AI-ingested documents cleaned up into reviewable source files and publishable output instead of leaving everything trapped in raw chat transcripts.

What it is: anthropics/defending-code-reference-harness is the Apache-2.0 reference implementation from Anthropic for autonomous vulnerability discovery and remediation with Claude, based on the team's learnings from partnering with security teams at several organizations since launching Claude Mythos Preview: the repository ships a Claude Code skill set

Why it matters: Useful for security teams and engineering teams that need an autonomous vulnerability discovery + remediation loop and want Anthropic's reference implementation as a starting point (not a vendor lock-in — the harness is configurable for any Claude API, including Bedrock, Vertex, and Azure): the reference pipeline runs in a gVisor sandbox by default so the autonomous patch step can execute target

What it is: bergside/design-md-chrome is an MIT Chrome extension that reads the active page's typography, colors, spacing, motion, and UI signals, then exports a reusable DESIGN.md or SKILL.md blueprint for Claude Code, Codex, Stitch, and similar AI design workflows.

Why it matters: Useful for designers and frontend builders who want to capture a live product's design system as structured AI-ready guidance instead of rebuilding the same visual tokens by hand.

What it is: Cloudflare's security-audit-skill is an MIT-licensed coding-agent skill that turns code review into a six-phase security audit pipeline with recon, parallel hunting, adversarial validation, machine-readable findings, and fresh-agent verification.

Why it matters: Useful for security-conscious teams that want agent-assisted audits to do more than dump a first-pass vulnerability list, especially when they need a repeatable process that tries to kill false positives before reporting them.

What it is: darkrishabh/agent-skills-eval is the MIT agent-skills-eval test runner from darkrishabh, the missing test framework for the Anthropic Agent Skills ecosystem. The motivation is sharp: Agent Skills make it easy to ship a SKILL.md and assume the agent is now better at the task — but the hard part is proving it. The architecture is paired-prompt evaluation with baseline subtraction: each

Why it matters: Useful for Agent Skills authors who want receipts — agent-skills-eval runs the same prompt twice (with_skill vs without_skill), has a judge model grade both, and produces a side-by-side HTML report so the skill author can prove the SKILL.md actually improves the model's performance rather than just adding noise. Useful for **Claude Code / Codex / OpenClaw / Hermes Agent skill library

What it is: google-deepmind/science-skills is the Apache-2.0 official Google DeepMind collection of agent skills for agentic scientific research workflows: each skill is a self-contained SKILL.md + scripts + references directory that an AI coding agent (Claude Code, Codex, Cursor, Gemini CLI, Antigravity, OpenCode, Hermes) can load to perform a specialized scientific task — AlphaGenome single-variant

Why it matters: Useful for AI research and engineering teams working on agentic scientific workflows: google-deepmind/science-skills is the official Google DeepMind collection of 30+ agent skills covering genomics, structural biology, cheminformatics, literature search, and clinical / pharmacological / regulatory databases, with every skill grounded in a primary scientific database

What it is: Google's Agent Skills repository packages reusable skill modules for Gemini, Agent Platform, Google Cloud, and related products so agents can start from product-specific procedures instead of generic prompt guesses.

Why it matters: That matters because a lot of agent work still breaks on vendor-specific setup details. An official skills repo gives builders a cleaner starting point when they want agents to work against real Google tools and cloud services.

What it is: apm microsoft/apm is an MIT-licensed open-source Agent Package Manager from Microsoft that treats MCP servers, agent skills, agent prompts, and reusable agent workflows as first-class installable artifacts (think npm for agents), with a registry format, versioned manifests, deterministic installs, and a single CLI that lets a coding agent (Claude Code, Codex, GitHub Copilot, etc.) discover

Why it matters: Useful for AI engineers, agent builders, platform teams, and coding-agent power users who want a package-manager-style install path for MCP servers and reusable agent skills instead of cloning dozens of repos or hand-curating ~/.claude/skills, because apm microsoft/apm from Microsoft ships a registry format, versioned manifests, deterministic installs, and a single CLI that integrates with

What it is: microsoft/skills collects official agent skill and MCP examples that help bridge Azure/Copilot-style patterns with practical agent integrations.

Why it matters: Useful for teams standardizing agent capabilities across prompts, services, and tool actions without inventing conventions for every app.

What it is: muratcankoylan/Agent-Skills-for-Context-Engineering is a 17,687-star, MIT-licensed comprehensive open collection of Agent Skills focused on context engineering and harness engineering principles for building production-grade AI agent systems, cited as foundational work in arXiv 2601.21557 'Meta Context Engineering via Agentic Skill Evolution'

Why it matters: Useful for AI agent builders, harness engineers, technical leads, and AI engineering teams who want a portable, open, MIT-licensed Skills pack that teaches context-engineering principles across any agent platform (Claude Code, Codex, Cursor, OpenClaw, Hermes), for AI teams that need an academic-backed reference for static skill architecture and self-improving contexts, for anyone shipping

What it is: Superpowers is an MIT-licensed agentic skills framework and software development methodology with marketplace installs for Claude Code, Codex App and CLI, Cursor, and other harnesses, plus reusable skills, onboarding docs, and harness-specific packaging.

Why it matters: Useful for teams that want a shared operating system for AI coding work instead of every developer reinventing skills, workflows, and review habits alone.

What it is: Protocol-Lattice/go-agent is an Apache-2.0, 252-star Go framework for AI agents with pluggable LLM providers (Gemini, OpenAI, Anthropic, Ollama, local dummy), short-term + vector-store-backed long-term memory, UTCP-native tool orchestration, CodeMode, sub-agent composition, deterministic guardrails, and a local .skills instruction loader that mirrors the SKILL.md layout.

Why it matters: Useful for Go engineers who want idiomatic agent runtimes without Python, for teams standardizing on UTCP for tool calls across stacks, and for builders who need pluggable providers + middleware (retry, timeout, rate, token budget) without leaving the Go ecosystem.

What it is: Agent Scan is an Apache-2.0 agent component inventory and threat scanner from Snyk that discovers installed MCP servers, harnesses, and agent skills on your machine, then scans them for prompt injections, sensitive-data handling, and malware payloads hidden in natural-language instructions. 2,587 stars, distributed as snyk-agent-scan on PyPI, supports Claude, Cursor, Windsurf, Gemini CLI, Amp

Why it matters: Useful for security teams that want one CLI to inventory and audit every MCP server, harness, and agent skill installed on a developer machine, and to flag prompt-injection or data-exfiltration patterns before they leak into production. Install via pip, run snyk-agent-scan against a config, and pipe the findings into your existing vulnerability dashboard.

What it is: TheLunarCompany/lunar is an MIT-licensed, open-source agent-native MCP gateway and registry from Lunar.dev that gives platform teams one governance, observability, and policy layer across every MCP server, agent skill, and tool an organization exposes, so they can run multi-vendor agent stacks at production scale without losing visibility or control.

Why it matters: Useful for AI platform engineers, agent operators, and security teams who need a single MCP gateway to govern, observe, and policy every MCP server, agent skill, and tool in an organization, so they can run multi-vendor agent stacks at production scale without losing visibility, audit trails, or the ability to revoke a single tool without breaking the whole agent fleet.