10 best Claude skills in 2026, compared by use case

The best Claude skills are Frontend Design for UI work, PDF for document jobs in Claude.ai, Skill Creator for reusable workflows, and Superpowers for end-to-end development. If a task needs repository access, Git, a CLI, or an authenticated MCP server, run the skill in Claude Code. For work contained in uploads, Project files, or connected services, Claude.ai may be enough.

This list favors a specific job, inspectable instructions or package details, visible dependencies, a concrete output, and limits you can check before installation. It includes official Anthropic skills, independent GitHub projects, and specialist marketplace listings.

The 10 best Claude skills at a glance

• Frontend Design: New or substantially redesigned web interfaces. Best in Claude Code with project files and development tools. Source: Anthropic GitHub.
• PDF: Creating, editing, extracting, and checking PDFs. Available as a built-in skill in Claude.ai and the API. Its GitHub folder is reference-only.
• Skill Creator: Building and evaluating custom skills. Works in Claude.ai and Claude Code. Source: Anthropic GitHub.
• Superpowers: A complete plan, build, test, and review workflow. Best in Claude Code. Source: community GitHub plugin.
• Grill With Docs: Pressure-testing plans before implementation. Best in Claude Code. Source: community GitHub bundle.
• Differential Review: Security-sensitive code review. Best in Claude Code. Source: Trail of Bits on GitHub.
• Hugging Face CLI: Model, dataset, job, and endpoint operations. Best in Claude Code with the hf CLI; private or write operations need credentials.
• Google Search Console Indexing: Diagnosing indexing problems. Works in Claude Code or from supplied exports. Marketplace listing; Search Console access is needed for live actions.
• TrustMRR Market Research: Revenue-backed competitor and market research. Best in Claude Code with TrustMRR MCP/API access. Marketplace listing.
• Web Performance Audit: Turning performance evidence into release decisions. Works in a compatible agent with supplied field, lab, trace, or source data. Marketplace listing.

What counts as a Claude skill

Anthropic defines Claude skills as folders of instructions, scripts, and resources that Claude loads when a task matches. Some work in Claude.ai. Others need Claude Code, a plugin, command-line tools, an MCP server, or API access.

Check that distinction before installing anything. A useful-looking SKILL.md can still fail if it assumes tools your Claude environment cannot access. For downloaded folders, read every bundled script before running it. The installation paths for Claude.ai and Claude Code are covered below.

Claude skills vs Claude Code skills

Custom skills can use the same format across Claude.ai and Claude Code, but they do not get the same working environment. A skill supplies instructions; the host controls which files, commands, services, and credentials it can actually use.

• Files: Claude.ai can use uploads, Project files, created files, and connected sources. Claude Code can use files in the local project and any additional directories you approve.
• Code: Claude.ai executes code inside an isolated container. Claude Code can run shell commands, scripts, builds, tests, and Git inside the working environment.
• External tools: Claude.ai can use enabled connectors, remote MCP servers, and permitted network access. Claude Code can use configured MCP servers plus installed and authenticated command-line tools.
• Best fit: Claude.ai suits documents, spreadsheets, presentations, PDFs, and bounded analysis. Claude Code suits repositories, automation, debugging, audits, deployments, and multi-tool workflows.

Claude.ai skills work inside the chat sandbox

In Claude.ai, skills use Anthropic's code execution and file creation environment. Claude can read chat uploads and Project files, use connected sources, run code in a private container, and return files it creates. That is a strong fit for the PDF skill, spreadsheet analysis, presentations, reports, and other jobs built around documents or supplied data.

Claude.ai can also use connectors and remote MCP servers. Its container may install approved packages or reach permitted domains when network access is enabled. Those settings vary by plan and organization. The chat does not automatically inherit your laptop's working tree, locally installed tools, shell configuration, or saved CLI credentials. A skill may contain a script or ask for a command, but that command must still exist and be allowed inside the container.

Claude Code skills can use the working environment

Claude Code runs in a terminal or IDE with access to the project you opened. With permission, it can read and edit repository files, search across directories, run shell commands, execute bundled scripts, build the app, run tests, inspect Git history, and check its own changes. This is why Frontend Design, Superpowers, Grill With Docs, and Differential Review are much more useful in Claude Code than in a standalone chat.

Claude Code can also combine project files with external tools. A Search Console skill can call an authenticated MCP server or a local CLI wrapper. TrustMRR can use its MCP connection. The Hugging Face skill can run the hf CLI with your configured account. The same applies to other installed CLIs, internal scripts, databases, and MCP servers. Access still depends on permissions and valid credentials, but the skill can coordinate those tools with the files and commands already available in the project.

A Claude Code skill can inspect a repository, query an external service, write a fix, run the relevant checks, and produce a report without moving every input through a chat upload. Use Claude.ai when the job is contained in documents, Project files, or connected sources. Use Claude Code when the workflow depends on local files, terminal commands, a codebase, or several external systems working together.

1. Best for frontend work: Frontend Design

Frontend Design is Anthropic's opinionated answer to generic AI-generated interfaces. It makes Claude establish a visual direction before coding, then forces deliberate choices about typography, color, layout, motion, copy, and the one detail that should make the page memorable.

Choose it for a new interface or a substantial redesign. For one component inside a settled design system, follow the existing system instead. Keep brand rules explicit because the skill favors aesthetic exploration.

2. Best for PDF work: PDF

Anthropic's PDF skill covers extraction, OCR, merging, splitting, rotation, forms, encryption, image extraction, and PDF creation. It includes practical library and command-line guidance rather than telling Claude to "work with this PDF" and hoping for the best.

This is the most broadly useful entry on the list for non-developers. It can help with reports, forms, scanned documents, and document cleanup in Claude.ai or through the API. Anthropic does not make this pre-built document skill available in Claude Code. Its GitHub folder is proprietary reference material, not an install target, and its license does not permit copying, derivatives, or redistribution outside the allowed Anthropic service use.

3. Best for creating custom skills: Skill Creator

Skill Creator is the right starting point when none of the public options fit your workflow. It helps define the trigger, expected output, folder structure, supporting resources, and evaluation prompts. More importantly, it treats the first draft as a draft.

Use it for workflows that repeat and have a visible quality bar. A weekly reporting process, release checklist, or research method is a better candidate than a one-off request. The evaluation loop takes more effort than asking Claude to save a prompt, but that effort is the point.

4. Best for end-to-end development: Superpowers

Superpowers is a plugin and skill bundle rather than one narrow skill. It covers brainstorming, planning, Git worktrees, test-driven development, debugging, code review, verification, and branch completion. The bundle is available through Claude Code's plugin system and supports several other coding agents.

Install it when you want Claude to follow a full engineering method instead of choosing individual skills yourself. Skip it if you prefer lightweight, manually invoked helpers. Superpowers is intentionally prescriptive, especially around planning and TDD.

5. Best for pressure-testing plans: Grill With Docs

Matt Pocock's Grill With Docs pushes Claude to question a proposed change before implementation. It also records the resulting domain language and architectural decisions instead of leaving the useful part trapped in chat history.

The skill is tiny because it composes other skills from the same repository. Install the full bundle and run its setup step rather than copying this one file in isolation. It is useful for ambiguous features, architecture changes, and projects where the same terminology keeps getting re-explained.

6. Best for security-sensitive reviews: Differential Review

Trail of Bits' Differential Review reviews code changes with a security lens. It uses git history, classifies risk, checks test coverage, estimates blast radius, and requires evidence for findings. High-risk changes can trigger adversarial analysis.

Authentication, cryptography, validation, external calls, and value-transfer code justify the overhead. Formatting and routine documentation changes do not. The workflow can run shell commands and inspect repository history, so review its permissions before adding it to a sensitive codebase.

7. Best for model and dataset work: Hugging Face CLI

The official Hugging Face CLI skill gives Claude a command reference for models, datasets, repositories, jobs, endpoints, caches, and authentication. It is useful when the task depends on exact hf commands rather than general machine-learning advice.

This is a fit only if your work already happens on Hugging Face. Some commands create, update, deploy, or delete remote resources, so start with read-only inspection and ask for a plan before authorizing mutations. You need the CLI; private or mutating operations also need credentials, and hosted jobs or endpoints may incur separate charges.

8. Best for indexing problems: Google Search Console Indexing

Google Search Console Indexing investigates excluded pages using URL Inspection, crawl evidence, HTTP status, robots rules, canonicals, redirects, sitemap membership, and internal discovery. It separates hard blockers from weaker quality or discovery hypotheses.

Choose it after a migration, template change, accidental noindex, or unexplained exclusion pattern. Live inspection or sitemap submission requires Search Console access; offline diagnosis can use supplied inspection exports and crawl evidence. It cannot guarantee indexing, and it does not treat repeated submission as a substitute for fixing the page.

9. Best for revenue-backed market research: TrustMRR Market Research

TrustMRR Market Research builds a bounded startup cohort, checks pagination and money units, separates payment-provider-backed observations from founder claims, and produces a decision memo with a short validation test. A bundled validator checks totals, quartiles, medians, duplicates, and concentration.

Use it for competitor screens, acquisition research, or an early market decision where observed revenue is useful evidence. TrustMRR is still an opt-in dataset, not a complete market map. The skill requires separate TrustMRR API access, an MCP client with secret-based authentication, and Python 3.

10. Best for web performance audits: Web Performance Audit

Web Performance Audit keeps field data, lab reports, traces, and source evidence separate before recommending fixes. It maps findings to routes and user journeys, then returns a release verdict, prioritized fixes, budgets, and a verification matrix.

Use it when a Lighthouse score changed, Core Web Vitals regressed, or a release needs a performance gate. It is designed to work from supplied evidence and includes a read-only report summarizer. If you have no comparable measurements, the honest output may be a measurement plan rather than a fix list.

How to install Claude skills

In Claude.ai, open Customize > Skills to browse available skills, connectors, and plugins. You can also upload a compatible custom skill. Code execution and file creation must be enabled for skills to run.

In Claude Code, standalone GitHub skills usually go into ~/.claude/skills/ for personal use or .claude/skills/ for one project. Plugin bundles use Claude Code's plugin flow instead. The Claude Code skill installation guide covers both scopes, verification, updates, and common folder mistakes.

For marketplace listings, check the compatible agents, required tools, package contents, and limitations before downloading. Search Console, TrustMRR, Hugging Face, and similar services still need their own access or credentials. A skill cannot create those permissions for you.

How to choose the right one

Pick the smallest skill that owns the expensive part of the job. For a single PDF, use the PDF skill. For a full engineering process, a bundle such as Superpowers may justify its extra machinery. For specialist work, check whether the package exposes its method, dependencies, examples, limitations, and expected output.

Before installing, ask:

• Does this solve a workflow you repeat?
• Does it work in the Claude surface you use?
• Can you inspect its instructions, scripts, dependencies, and license?
• Does it state which tools and credentials it needs?
• Can you verify the output instead of trusting a polished answer?

Free GitHub skills are often the best answer for common workflows. Paid skills need to earn their price with deeper method, supporting files, evaluation cases, maintenance, or reduced inspection time. Source and price are useful filters. Neither proves quality on its own.