Best AI Agent Skills

A deep look at nextlevelbuilder's open-source design-intelligence skill: 84 UI styles, 192 color palettes, 74 font pairings, and 161 industry reasoning rules that make a coding agent commit to a design system before it writes any UI code.

Anthropic's own open-source Agent Skills repo: the SKILL.md spec plus reference skills for docx, pptx, xlsx, and PDF work. The foundation every other skill on this list, and most of the ecosystem, is built on top of.

Jesse Vincent's (obra) skill-driven development methodology for coding agents: mandatory brainstorming before code, two-to-five-minute task plans, strict red-green-refactor TDD, and subagent-reviewed commits. The single most-cited "this changed how I work with Claude Code" repo of 2026.

davila7's npx-installable CLI and live dashboard for Claude Code: a searchable catalog of 900+ agents, commands, hooks, MCPs, and skills, plus real-time session analytics and health checks, all installable without editing a single config file by hand.

A multi-harness plugin marketplace of 94 plugins, 203 subagents, 175 skills, and 109 commands, written once and shipped natively to Claude Code, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, and OpenCode instead of a lowest-common-denominator export.

punkpeye's curated directory of hundreds of Model Context Protocol servers: the layer that lets Claude Code read a database, query Jira, or drive a browser directly instead of working from copy-pasted context. This is where you go before you build a connector yourself.

Corey Haines' open-source marketing skill pack: CRO, copywriting, cold email, SEO audits, ad creative, analytics, churn prevention, and growth engineering, organized as one skill per task so Claude Code can run a real marketing operation instead of just drafting copy.

A universal SEO plugin for Claude Code: 25 sub-skills and 18 parallel specialist agents covering technical SEO, E-E-A-T, schema markup, AI search (GEO/AEO), backlinks, and local SEO, producing prioritized, falsifiable action plans instead of a generic checklist.

alirezarezvani's 362-skill library spanning engineering, marketing, product, and (the part that stands out) 68 C-level advisory skills, 19 compliance/QM skills, and dedicated finance and commercial-operations packs, all running on Python stdlib tooling with zero pip installs.

50 skills distilled from specific bestselling business and marketing books (Blue Ocean Strategy, StoryBrand, $100M Offers, Hooked, Made to Stick) plus 12 "metaskills" that chain them into a guided journey like "grow this business" instead of leaving you to pick a framework yourself.

120 marketing skills across 7 disciplines (SEO/GEO, paid ads, email, social, influencer, launch, brand narrative): the distinguishing feature is 8 auditor skills that score every output against a named framework (ROAS, SEND, CORE-EEAT) and hand back a typed SHIP / FIX / BLOCK verdict instead of just prose.

A paid-media operations skill for Claude Code spanning 12 ad platforms (Google, Meta, YouTube, TikTok, LinkedIn, and more) with source-grounded audits, deterministic scoring, and versioned JSON reports. Read-only by default; account changes require explicit capability grants.

Ads Skills packages LinkedIn, Meta, and Google Ads playbooks (40+ strategy files, 39 API scripts) from ColdIQ's Head of Growth, who runs $200K+/month in real B2B ad spend. See what's inside and how to install it.

AI Ads for Claude Code turns a business website into a full ad strategy: platform-specific copy for Google, Meta, LinkedIn, TikTok, YouTube, and Pinterest, funnel design, budget splits, and a PDF report, with no ad account logins or API keys required.

GitHub's official spec-driven development toolkit: 125k+ stars, MIT-licensed, working with 30+ AI coding agents. Instead of vibe-coding from a one-line prompt, you build a constitution, spec, plan, and task list the agent executes step by step: /speckit.specify → /speckit.plan → /speckit.tasks → /speckit.implement.

Microsoft's official MCP server that gives coding agents a real browser: 35k+ stars, Apache 2.0. It drives pages through Playwright's accessibility tree instead of screenshots, so Claude can click, type, fill forms, and verify your app deterministically. The single most-installed testing tool in the MCP ecosystem.

Upstash's MCP server that pipes up-to-date, version-specific library documentation straight into your agent's context: 60k+ stars, MIT. Ends the "Claude hallucinated an API that doesn't exist" problem by fetching real docs and code examples from source at prompt time. Append "use context7" and outdated training data stops mattering.

Supabase's official MCP server: your agent queries the database, runs migrations, reads logs, checks security advisors, deploys Edge Functions, and generates TypeScript types directly. Hosted with browser OAuth (no token pasting), with read-only mode and per-project scoping as first-class security controls.

Anthropic's official AI security reviewer: a GitHub Action that scans every pull request for real vulnerabilities (injection, auth bypasses, hardcoded secrets, unsafe deserialization) and comments on the exact lines, plus a /security-review command for Claude Code. Semantic analysis with aggressive false-positive filtering, in any language.

fourteenwm's open-source Claude Code skill pack for Google Ads specialists: 42 skills across 9 categories, from search query mining to RSA testing to portfolio bids, built with real guardrails against accidental account mutations by someone running 110+ live accounts.