What Are Claude Skills and Why Do They Matter?
A Claude skill is a SKILL.md file that tells Claude Code how to behave inside a specific project. It stores conventions, build steps, naming patterns, and “we do not do this because of that incident” knowledge. Without skills, Claude starts every session cold and fills gaps with confident guesses. With skills, the agent reuses accumulated intent.
Skills are the difference between a generic assistant and a project-native engineer. They are especially important in 2026 because agentic workflows are now the default. A loop that runs unattended needs to know the rules without asking.
The Top Claude Skills of 2026
Based on install data from the skills directory, the most-used skills cluster around frontend engineering, code review, cloud platforms, and agentic tooling. Here is the ranking.
| Rank | Skill | Publisher | Installs | Purpose |
|---|---|---|---|---|
| 1 | find-skills | Vercel Labs | 2.3M | Discovers and installs relevant skills for the current project |
| 2 | frontend-design | Anthropic | 625.1K | Design guidelines and component patterns for UI work |
| 3 | vercel-react-best-practices | Vercel Labs | 525.4K | React conventions for Next.js and Vercel projects |
| 4 | agent-browser | Vercel Labs | 511.3K | Browser automation and web agent workflows |
| 5 | grill-me | Matt Pocock | 454.0K | Interactive questioning to clarify and improve code |
| 6 | web-design-guidelines | Vercel Labs | 438.1K | General web design principles for agentic projects |
| 7+ | Microsoft Azure skills | Microsoft | 6.8M total | Cloud-specific skills for Azure, compute, migration, and Foundry |
| 23 | remotion-best-practices | Remotion | 408.2K | Video generation and editing with React |
| 24 | azure-hosted-copilot-sdk | Microsoft | 400.0K | Building Azure-hosted copilot applications |
| 25 | grill-with-docs | Matt Pocock | 375.2K | Code review against project documentation |
| 26 | improve-codebase-architecture | Matt Pocock | 372.6K | Refactoring and architectural improvement prompts |
| 27 | azure-compute | Microsoft | 370.7K | Azure compute resource management |
| 28 | azure-cloud-migrate | Microsoft | 360.7K | Cloud migration guidance for Azure |
The numbers tell a clear story: the most popular skills are not exotic. They are everyday engineering conventions packaged so Claude can follow them without being reminded every session.
Skill Categories That Dominate
1. Frontend and Design
frontend-design, web-design-guidelines, and vercel-react-best-practices are near the top. This makes sense. UI work is high-volume, opinionated, and full of project-specific patterns. A skill that encodes the design system, component structure, and state management rules saves a huge amount of back-and-forth.
2. Code Review and Refinement
Matt Pocock’s grill-me, grill-with-docs, and improve-codebase-architecture skills are widely used because they structure the review process. Instead of asking “is this code good?” you get a focused interrogation against the codebase and docs. This is the kind of adversarial review that catches the mistakes the implementing agent misses.
3. Cloud and Enterprise
Microsoft’s Azure skills collectively dominate the install counts. This reflects the reality that enterprise AI work is happening inside cloud platforms. Skills for compute, migration, Foundry, and hosted copilots mean Claude can operate within the constraints of Azure without the engineer manually pasting documentation every time.
4. Agentic and Browser Automation
agent-browser and find-skills point to the agentic shift. The browser skill lets Claude navigate and interact with web interfaces. The skill finder lets projects auto-discover what they need. These are meta-skills for the loop engineering era.
How to Use a Skill in Claude Code
Skills are invoked by name or matched implicitly based on the task. A typical workflow:
# Install a skill into your project
claude skills add vercel-react-best-practices
# In a prompt, invoke the skill explicitly
$vercel-react-best-practices
Review this Next.js page and apply the project conventions.
Or the skill can be configured to activate automatically when the task matches its description. The tighter the skill description, the better the auto-match.
A minimal skill file looks like this:
# Skill: React Best Practices
## Conventions
- Use Server Components by default.
- Use Client Components only when interactivity is required.
- Prefer `fetch` with explicit caching headers.
- Colocate tests next to the component in `__tests__/`.
## Anti-patterns
- Do not use `useEffect` for data fetching.
- Do not import heavy libraries into Client Components without lazy loading.
This is the entire trick. The skill is a small, durable file that replaces repeated explanations.
How to Build Your Own Skill
If you find yourself explaining the same thing to Claude more than twice, turn it into a skill.
- Identify the repeated context. Common build commands, test patterns, naming conventions, security rules.
- Write the skill file. Keep it short. Use bullets, examples, and anti-patterns.
- Test it. Run a task with the skill and verify the output follows the conventions.
- Iterate. Add rules as the agent makes new mistakes.
- Share it. Package it as a plugin if the same patterns apply across multiple repos.
Skills vs. Other Steering Mechanisms
| Mechanism | Best For | Scope |
|---|---|---|
CLAUDE.md | Project-wide constitution and high-level rules | Entire repo |
SKILL.md | Specific domain or workflow conventions | Named skill, reusable across tasks |
AGENTS.md | Multi-agent definitions and team structure | Agent teams |
| System prompt | One-time behavior shaping | Single session |
Use CLAUDE.md for the rules that never change. Use SKILL.md for the conventions that apply to specific domains. Use AGENTS.md when you have subagents with different roles.
What I Learned About Conventions at Google
At Google, large codebases survive because conventions are written down and enforced. Style guides, build rules, and design patterns are documented so engineers do not have to relearn the system every time they touch a file. The same principle applies to Claude. The agent is faster, but it is still starting from zero unless you give it the conventions.
Skills are the convention system for AI agents. The teams that write them will move faster because the agent will stop guessing.
Frequently Asked Questions
Where do I find skills?
The skills directory at skills.sh lists published skills. You can also create your own for internal use.
Are skills free?
Many are open source. Some publishers may offer premium skills. Read the skill’s license before using it commercially.
Can a skill hurt my codebase?
A bad skill can. A skill that encodes outdated or wrong conventions will make the agent repeat those mistakes. Treat skills as code: review them, test them, and version them.
Do I need skills for every project?
No. Small, throwaway projects do not need them. Projects with multiple agents, long-lived code, or team handoffs benefit the most.
What is the best skill to start with?
find-skills is a good default. It helps the project discover the skills it needs. From there, add a skill that matches your stack.
References & Further Reading
- skills.sh — directory of Claude and Codex skills
- Claude Code Documentation — official guide to skills
- Anthropic Claude Code — terminal-based agent that uses skills
- OpenAI Codex — agentic coding assistant with similar skill support
- Vercel AI SDK — tools for building agentic and frontend AI applications
What to Read Next
Skills are the building blocks of reliable agentic workflows. Once you have skills, the next step is combining them into loops that run without you driving every turn.
- 📖 Related: Loop Engineering with AI: How to Build Self-Running Agents That Ship
- 📬 Newsletter: Get AI engineering posts like this weekly → Subscribe free
- 📥 Resource: The AI Mastery Roadmap for Software Engineers