Author: Daniel Slug: top-claude-skills-ai-agent-developers-2026 Meta Description: We ranked the top Claude Skills for AI agent developers by GitHub stars, maintenance activity, and real-world fit. Updated Q2 2026 with data from 24 repos. Primary Keyword: claude skills for ai agents Secondary Keywords: best claude skills, claude code skills ranking, ai agent skills, claude skill marketplace Reading Time: 13 minutes Cluster Role: Pillar Pillar Parent: — Cover Image: ![Top Claude Skills for AI Agen

Guess how many Claude Skills are now indexed across the public marketplaces?

Over 5,000 on VoltAgent's awesome-list alone. Another 125,000+ stars sitting on the official anthropics/skills repo. 172,000+ on Obra's superpowers library. SkillsMP claims 900,000+. And the number keeps compounding every week.

For an AI agent developer, that's equal parts blessing and problem. Progressive disclosure means Claude can load dozens of skills without blowing the context window — but picking which dozens, out of five-figure options, is now its own research project.

This article is the research project. We pulled public GitHub data on 24 of the most-cited Claude Skill repositories across eight marketplaces — skills.sh, skillsmp.com, clawhub.ai, aitmpl.com, agent-skill.co, travisvn/awesome-claude-skills, heilcheng/awesome-agent-skills, and VoltAgent/awesome-openclaw-skills — and ranked the ten most relevant to developers building AI agents that touch the web, the shell, or both.

📮 What we found: The top three skills hold ~99% of all stars in this category. Everything below 2,500 stars is still finding its audience. And every browser-automation skill on the list runs vanilla Playwright — none ship with stealth defaults. That last gap, more than anything else, is where 2026 will play out.

How We Ranked These Skills

A top-ten listicle without a scoring method is a vibes document. Here is ours.

We used three weighted signals:

• GitHub stars (50%) — the strongest public proxy for voluntary adoption. Skills with thousands of stars have been audited, tested, and linked from more places than Anthropic itself can enumerate.
• Maintenance activity (30%) — last-commit date and release cadence. An abandoned skill with 5,000 stars is worse than an active one with 500.
• Agent-developer fit (20%) — subjective scoring from the authors, based on how tightly the skill helps a developer build an agent (versus consume one). A business-analytics skill is excellent but not on this list.

We excluded skills that have no public GitHub presence (community skills that exist only as clawskills.sh listings, for example) because their install counts are not publicly verifiable. That's a deliberate constraint: this ranking favors things you can read, fork, and audit.

Every star count below was pulled on April 30, 2026 and is accurate as of that date. The Anthropic-official entries share one parent repository (anthropics/skills), so they show the same star count — we still list them separately because Claude treats them as distinct skills at runtime.

The Top 10 Claude Skills for AI Agent Developers

1. obra/superpowers — 172,382 ⭐

If you've used Claude Code for more than a week, you have probably already encountered Superpowers. Built by Jesse Vincent (obra), it's a collection of 20+ battle-tested skills covering test-driven development, debugging, pair-programming patterns, /brainstorm and /write-plan slash commands, and a built-in skills-search tool that finds other installed skills for you.

Nothing else in the Claude Skills ecosystem is close on adoption. The repository passed 100,000 stars in late 2025 and has kept compounding — it is the de facto "standard library" of community skills.

Why developers install it:

• Battle-tested workflow primitives — /brainstorm, /write-plan, /execute-plan cover the planning phase end-to-end without bespoke prompts.
• Skills-search built in — you don't have to remember which skill does what; the bundle indexes itself.
• 20+ skills, one install — TDD, debugging, code review, refactoring patterns all bundled, all auditable.

The honest tradeoff: Superpowers is a kitchen sink. The individual skills are strong, but discovery inside the bundle is not great — you install it, then learn what's there over weeks. Treat it as a base layer, not a final answer.

💡 Pro Tip: Pair Superpowers with one specialty skill (browser, scraping, MCP) rather than stacking three kitchen-sink bundles. Command ambiguity is the #1 reason teams uninstall after a week.

Install: /plugin marketplace add obra/superpowers-marketplace
Repository: obra/superpowers

2. anthropics/skills — mcp-builder — 125,856 ⭐

This is Anthropic's official skill for scaffolding a high-quality MCP (Model Context Protocol) server. Hand it an API description and it produces a well-structured MCP server you can wire into any Claude client — Claude.ai, Claude Code, the API, or a custom agent.

It earns its #2 spot because MCP is the integration layer almost every non-trivial agent eventually needs, and because this is the official builder maintained by the same team shipping Claude itself. Reading the generated output is a free tutorial in Anthropic's preferred skill structure.

Why developers install it:

• Official conventions encoded — frontmatter, folder layout, naming all match Anthropic's spec by default.
• API-to-MCP shortcut — describe an external API, get a working server back in minutes instead of hours.
• Free MCP onboarding — reading the scaffold output is the fastest way to learn the spec.

The honest tradeoff: The output still requires you to understand authentication, rate limits, and the MCP spec itself. It's a scaffold, not a no-code generator. Plan on a half-day of integration work after the scaffold drops.

Repository: anthropics/skills — mcp-builder

3. anthropics/skills — webapp-testing — 125,856 ⭐

A Playwright-based skill for testing local web applications. Ship it a running dev server and it spins up a headless browser to verify behavior, capture screenshots, and debug UI regressions — all driven by Claude's reasoning rather than hand-written test scripts.

It's the third-most-installed Anthropic-official skill for one simple reason: the feedback loop is tight. You ask Claude to add a feature, it writes the code, then it tests the feature end-to-end in a real browser before telling you it's done.

Why developers install it:

• Real-browser verification — closes the "agent wrote the code but did the UI actually work?" gap that type-checking can't cover.
• Auto-screenshot on failure — bug reports come with visual evidence by default.
• Selector-free testing — Claude reasons about the page, so you don't maintain brittle CSS selectors.

The honest tradeoff: This is Playwright, not a stealth browser. Any production site with Cloudflare, DataDome, or PerimeterX in front of it will block this skill on request one. Fine for staging and internal apps; not fine for testing flows that touch real third-party platforms — see the gap analysis below.

👀 Did You Know? The first time most teams hit this wall isn't on a production target — it's a third-party SSO redirect. The login flow your QA suite has tested against staging for six months breaks the moment it points at the real OAuth provider.

Repository: anthropics/skills — webapp-testing

4. SawyerHood/dev-browser — 6,027 ⭐

A Claude Skill whose one-line description is exactly what it sounds like: "give your agent the ability to use a web browser." Single-focus, well-documented, fast install.

Among community-published browser skills, this has by a wide margin the most stars — an order of magnitude above the next entry in the niche. Simplicity and a clean README matter.

Why developers install it:

• Smaller surface area than webapp-testing — easier to reason about when your agent only occasionally browses.
• Clear README, fast onboarding — the install-to-first-use time is under five minutes.
• General-purpose, not test-specific — works for browsing, scraping known pages, or quick page-state checks.

The honest tradeoff: Same as #3 — vanilla Playwright under the hood. Cloudflare, DataDome, and PerimeterX will all flag it quickly on targets that defend against automation.

Repository: SawyerHood/dev-browser

5. lackeyjb/playwright-skill — 2,518 ⭐

A general-purpose Playwright wrapper packaged as a skill. Less opinionated than dev-browser, more flexible for custom automation scripts, and the most-starred direct bridge to Playwright in the Claude Skills ecosystem.

For teams with existing Playwright test suites, integration cost is close to zero — you keep the suite, you add Claude as a driver.

Why developers install it:

• Drop-in for existing Playwright code — your selectors, your locators, your fixtures all keep working.
• Less prescriptive — no built-in test-runner conventions; you bring your own.
• Maximum control — closest you can get to "Claude calls Playwright APIs directly" while still being a skill.

The honest tradeoff: If you don't already know Playwright, pick webapp-testing or dev-browser first. This skill assumes the mental model — locators, page objects, async patterns — is already in your head.

Repository: lackeyjb/playwright-skill

6. browserwing/browserwing — 1,253 ⭐

A hybrid tool that exposes browser actions as either MCP commands or a Claude Skill. The architectural pitch is token efficiency — native commands instead of LLM-driven page reasoning for deterministic actions like "click this known button" or "fill this known form."

It's the most visible entry in the "MCP-first browser automation" subcategory, and the category is growing faster than star counts suggest.

Why developers install it:

• Token budget control — deterministic actions don't burn frontier-model tokens for no reason.
• MCP-native — slots into existing MCP-server stacks without an extra adapter layer.
• Aggressive release cadence — one of the most actively maintained mid-size skills in this list.

The honest tradeoff: Newer project, smaller community than the top five. Migration away is easy; relying on it for production requires pinning a version and watching the changelog.

Repository: browserwing/browserwing

7. anthropics/skills — skill-creator — 125,856 ⭐

An interactive skill that guides you through creating your own skills — ask questions, generate the SKILL.md file and folder structure, test locally, ship.

The moment you notice you're prompting Claude the same way in three different projects, you have a skill on your hands. This is the official way to formalize it.

Why developers install it:

• Valid frontmatter on the first try — no learning the YAML schema by trial and error.
• Generates folder structure too — scripts/ and resources/ subdirs come with the scaffold.
• Q&A interview format — you describe the workflow in plain English; the scaffold matches.

The honest tradeoff: The scaffold is basic. For anything beyond a single-prompt skill, you'll rewrite ~50% of the generated SKILL.md to add the nuance the interview can't capture.

Repository: anthropics/skills — skill-creator

8. firecrawl/cli — 353 ⭐

Four official skills from Firecrawl — firecrawl-cli, firecrawl-agent, firecrawl-scrape, and firecrawl-search — that expose their hosted scraping, crawling, search, and autonomous-agent capabilities to Claude.

The skill-layer star count (353) is small. The parent project (firecrawl/firecrawl) sits at 112,988 stars. That two-orders-of-magnitude gap is the most interesting data point in this whole ranking — the underlying tool is wildly adopted; the skill layer hasn't been claimed yet.

Why developers install it:

• Hosted backend — the scraping itself runs on Firecrawl's infrastructure, not yours.
• Four skills, one install — scrape, crawl, search, autonomous-agent all wired in.
• Production-grade rate handling — the Firecrawl API does retry and backoff for you.

The honest tradeoff: Firecrawl is a hosted service with its own pricing. The skill is free; the backend is not. For simple one-off scrapes, dev-browser plus a Playwright selector is cheaper.

💡 Pro Tip: Watch this skill's star count over the next two quarters. If it crosses 2,000, the scraping skill category has consolidated. If it doesn't, the window for new entrants is still open.

Repository: firecrawl/cli

9. mhattingpete/claude-skills-marketplace — 563 ⭐

A marketplace-style bundle of Claude Code skills for software-engineering workflows — Git automation, test-runner orchestration, code review, release management. Less ambitious than Superpowers, sharper on the specific SWE primitives.

It earns its spot because of the specificity. Where Superpowers gives you 20 general primitives, this gives you 8-10 sharp ones aimed at "agent runs the SWE loop end-to-end."

Why developers install it:

• Sharper SWE primitives — Git, tests, review have explicit slash commands.
• Better discoverability than kitchen-sink bundles — fewer skills, clearer names.
• Drop-in CI fit — most of the skills work as well in headless CI as in interactive Claude Code.

The honest tradeoff: Some overlap with Superpowers' native commands. Pick one as your primary SWE-workflow layer to avoid command ambiguity; running both leads to "which /review are we calling?" debugging.

Repository: mhattingpete/claude-skills-marketplace

10. heal-dev/heal-playwright-tracer — 32 ⭐

An open-source statement-level Playwright tracer purpose-built for AI agents. Produces structured trace output that an LLM can reason about — turning "test failed" into "test failed because the click on selector X resolved to a hidden element after step Y."

Why does a 32-star skill earn the #10 spot? Because it solves a specific, real problem nobody else solves directly. Low-star, high-uniqueness skills earn ranking spots when they fill a gap.

Why developers install it:

• Structured failure analysis — Claude can read the trace and explain what happened.
• Statement-level granularity — pinpoints the exact line, not just the test name.
• Purpose-built for agent reasoning — output format is LLM-friendly by design.

The honest tradeoff: Small project, small contributor base. If you depend on it in production, be prepared to read the source — and consider pinning a version.

Repository: heal-dev/heal-playwright-tracer

At-a-Glance Comparison

Three observations jump out from this table.

First, the top three are all from Anthropic or obra. The gap between the official-plus-obra tier and everything else is two orders of magnitude wide. For a new agent developer, starting with those three covers maybe 70% of what you'll need on day one.

Second, every browser-automation skill on this list uses vanilla Playwright. None of them ship with stealth defaults, residential proxies, or captcha-solving. That is the single largest gap in the Claude Skills ecosystem as of April 2026.

Third, scraping-specific tooling is underrepresented. Firecrawl's skill layer is new (353 stars) despite the parent project having 113,000. Other specialized scrapers — for Amazon, X/Twitter, LinkedIn — exist on community marketplaces but lack the open GitHub presence that makes them auditable.

The Gap: What These Skills Don't Cover

If you are building an agent that needs to interact with real production websites — the kind with Cloudflare challenges, rotating CAPTCHAs, and aggressive fingerprinting — none of the top ten will carry you all the way.

The failure mode is predictable. You start with webapp-testing or dev-browser, your agent works beautifully against httpbin.org and your own staging environment, and then the first time it hits a real target — a SERP page, a LinkedIn profile, a marketplace listing — it gets a captcha on request one and a 403 on request two. The skills are not wrong; they just weren't built to defeat commercial anti-bot systems.

🤯 The numbers worth remembering: Firecrawl parent project — 112,988 stars. Firecrawl skill layer — 353 stars. The most-starred community browser skill — 6,027. There is no public Claude Skill in 2026 that combines stealth browsing with structured scraping. That slot is still empty.

This is where specialized browser infrastructure starts to matter. BrowserAct is purpose-built for this exact gap — stealth mode, residential proxies, automatic CAPTCHA bypass, and structured data extraction. It's not on this ranking because it currently exposes a REST API and templates rather than a Claude Skill, but agent developers integrating it via an MCP layer report the same pattern: the top-ten skills handle the coding and orchestration, and BrowserAct handles the "getting past the bouncer" part of the job.

For contextual work — like monitoring product listings — developers pair it with the Amazon Product API skill or the Reddit Posts & Comments Scraper template rather than reinventing selectors from scratch.

We expect the stealth-browser category to have a dedicated, well-starred Claude Skill within 2026. If you are a skill author looking for a gap with real demand, this is it.

Who Should Install What

If you are just starting with Claude Code for agent development, here is the minimum viable stack:

• obra/superpowers as your base SWE layer.
• mcp-builder the first time you need a new integration.
• webapp-testing or dev-browser when you need browser access.
• skill-creator the first time you catch yourself repeating a prompt.

For a team building an agent that browses real production sites:

• All four above.
• firecrawl/cli for structured scraping.
• browserwing if token budget is a pressure point.
• A stealth-browser backend (BrowserAct, Browserbase, or similar) behind an MCP server — no community skill covers this yet.

For a team building a developer-tool agent that lives inside repositories:

• obra/superpowers plus mhattingpete/claude-skills-marketplace — some overlap, but the SWE-specific one has sharper primitives for Git, tests, and reviews.
• skill-creator to codify team-specific workflows.
• heal-playwright-tracer if your agent runs browser tests.

Conclusion

The skill market looks crowded at 5,000 entries, but the productive surface is smaller. Ten skills, ranked by measurable adoption, cover the majority of what an AI agent developer actually needs — coding harness, MCP integration, browser access, scraping, skill authoring, and test analysis. Everything else is either specialization, redundancy, or still early.

The most interesting signal is where the ranking thins out. The best browser skill has 6,000 stars. The best scraping skill has 353. A stealth-browser skill that combines both capabilities cleanly does not yet exist in public form — which, for agent developers who need to work against real production sites, remains the largest gap between what Claude Skills promise and what they currently deliver.

If you're hitting that wall, take a look at BrowserAct — it's the infrastructure layer built to sit behind your agents at exactly this step, and the one that will most likely appear as a first-class Claude Skill in the near future.