Claude Code Skills 2026: The 10 You Actually Need

TL;DR: 500+ skills exist¹. Most add noise. These 10 earned permanent global install after 6 months of daily use² across 170+ pages of production code³: vercel-react-best-practices, next-best-practices, shadcn-ui, seo-content-writer, seo-fundamentals, tailwind-css-patterns, web-design-guidelines, architecture-patterns, clean-code, and fixing-accessibility. Plus how to build your own in 20 minutes.

What Skills Are

Skills are folders containing a SKILL.md file with YAML frontmatter (name and description) plus markdown operational rules and optional helper scripts. Claude preloads the short YAML description at startup and reads the full SKILL.md body when your prompt matches the task.

The vercel-react-best-practices skill contains 40+ rules about React inside Vercel — avoiding unnecessary useEffect, preferring Server Components, structuring data fetching. Claude reads those rules and applies them as constraints.

Skills vs Plugins vs Commands vs Subagents

Skills hold the reusable standards. Slash commands invoke repeatable workflows. Plugins distribute everything across VS Code, JetBrains, and CLI setups — see Claude Code plugins vs skills and the interactive Skills Hub.

The 10 Essential Skills

1. vercel-react-best-practices

40+ rules for React in Vercel apps. Every component starts Server-first. Shifts to Client only when interaction requires it. Saves bundle weight across 170+ pages.

2. next-best-practices

App Router fundamentals. Layout nesting, route groups, parallel routes, metadata API. Correct route structure on the first try.

3. shadcn-ui

shadcn/ui integration patterns. Extends through cn() and variant props. Keeps the design system coherent instead of fighting internal styles.

4. seo-content-writer

Content optimization for long-form writing. TL;DR-first structure, question-based H2s, FAQ inclusion, internal linking rules.

5. seo-fundamentals

Core SEO — crawlability, title tags, meta descriptions, canonicals, structured data. Caught 3 missing canonicals and 2 duplicate titles before I ran a crawl.

6. tailwind-css-patterns

Utility-first CSS conventions. Class grouping: layout → spacing → typography → color → state. Every component readable at a glance.

7. web-design-guidelines

Visual design for web. Spacing systems, typographic hierarchy, color contrast, visual weight. Shifts Claude from "technically correct" to "visually intentional."

8. architecture-patterns

System design for production apps. Separation of concerns, data flow, service boundaries. Kept the ACOS architecture clean as it grew.

9. clean-code

Code quality rules. Naming conventions, function length, abstraction levels. Every function name becomes a description of what it does. Code reviews are faster.

10. fixing-accessibility

WCAG 2.2 compliance. Semantic HTML, ARIA labels, keyboard navigation, focus management, color contrast. Passed WCAG 2.2 AA on first audit for the prompt library.

How to Install

# Install globally
npx skills add vercel-labs/agent-skills -y -g
npx skills add vercel-labs/next-skills -y -g
npx skills add sickn33/antigravity-awesome-skills -y -g
npx skills add ibelick/ui-skills -y -g
npx skills add giuseppe-trisciuoglio/developer-kit -y -g

# List installed
npx skills list -g

Building Your Own (20 minutes)

Create ~/.agents/skills/my-skill/SKILL.md:

---
name: my-brand-voice
description: Brand voice rules for content
triggers:
  - blog post
  - landing page
  - copy
---

# Brand Voice

## Rules

- Lead with outcomes, explain mechanisms second.
- Specific numbers over vague claims.
- Short sentences for impact. Longer for explanation.

## Patterns to Avoid

- Filler phrases: "it is worth noting", "in today's world"
- Passive constructions when active is available.

Full skill-building guide: Building Custom Skills for ACOS and starter repository configs in the Claude Code templates starter kit.

Evaluating Skills Before Global Install

A prose skill earns its keep only when you verify both trigger accuracy and output compliance. Before adding any skill globally:

1. Test positive and negative triggers — Confirm the YAML description fires on target tasks without hijacking unrelated prompts.
2. Pair prose with deterministic scripts — Put link checks, schema checks, and AST checks inside scripts/ so Node or Python executes exact assertions.
3. Install from audited MIT sources — Browse the 114-skill open catalog in frankxai/claude-skills-library or generate your own SKILL.md in the Skill Builder.

The ACOS Layer

The 10 skills above are the general foundation. ACOS adds 75+ domain-specific skills⁴: agent profiles, command libraries, hook systems, and memory architecture. The difference between giving Claude a style guide and giving Claude a complete operating context. Full setup: ACOS + Claude Code + MCP guide.

FAQ

How many skills globally?

Start with 5-10. More adds context overhead. I run 12 globally, load more per project.

Do skills work on Claude.ai?

Skills as filesystem files are Claude Code-specific. Claude.ai uses system prompts — same principles, different mechanism.

Can skills conflict?

Yes. Two skills with opposing rules for the same situation create inconsistent behavior. Review for overlap before combining.

How do I verify a skill loaded?

Ask Claude: "What skills are active?" It lists loaded context rules.

Global vs project-level?

Global (-g): universal standards. Project-level: domain-specific rules. Use both.