Best Claude Code skills in 2026

Below are 15 skills and plugins grouped by task. Each has a link to its catalog card, its access level to your project and one install command.

Which Claude Code skills should you install first?

These five cover the most common requests to the agent and barely overlap.

1. Superpowers sets a workflow from questions to review.
2. Karpathy-inspired guidelines keep the agent from writing extra code and making extra edits.
3. Planning with Files keeps the task plan in project files.
4. Anthropic's skills handle Word, PDF, PowerPoint and Excel.
5. The Codex plugin sends changes to a second model for review.

Access levels are given in words. Minimal means the skill is instructions only. Extended means it runs scripts, adds hooks or sends data to an outside service. Broad goes to skills that record your session history, register many hooks or download executables.

Skills for the development workflow

Superpowers

Superpowers is a plugin with skills that a whole task runs through. The agent first works out what you want and shows the design in parts for approval. Then it writes a plan of short tasks, each with file paths and a way to verify it, and hands them to subagents with two-stage review. The test-driven-development, systematic-debugging and using-git-worktrees skills switch on by themselves when the situation fits.

Access is extended. The plugin adds a session-start hook, and subagents run commands in your project for long stretches without you. The process applies even to a one-line fix, and the TDD skill deletes code written before its tests.

/plugin install superpowers@claude-plugins-official

Karpathy-inspired guidelines

The Andrej Karpathy-inspired guidelines fit in one file with four principles. The agent states its assumptions and asks when a task is unclear. It writes exactly the code that was requested, changes only the lines it needs to and defines up front what counts as done. There are no hooks or checks, so the result depends on how well the model sticks to the rules. Access is minimal.

npx skillfoxx add skills/karpathy-inspired-claude-code-guidelines

Matt Pocock's skills

Matt Pocock's skills are a few dozen small skills you call one at a time. grill-with-docs questions you before a change, writes shared project terms to CONTEXT.md and decisions to ADRs. tdd runs the loop of a failing test, working code and refactoring, and diagnosing-bugs walks through debugging step by step. Others cover specs, tickets, merge conflicts and handing context to another agent.

Run /setup-matt-pocock-skills before the first use in a repository. Access is minimal, though some skills create commits and tracker issues.

npx skillfoxx add skills/skills-for-real-engineers

Addy Osmani's Agent Skills

Agent Skills split development into stages, each with its own slash command, among them /spec, /plan, /build, /test, /review and /ship. Behind them are 25 skills with steps, checkpoints and a table of excuses the agent uses to skip a step. The author based it on the book Software Engineering at Google. Access is minimal.

npx skillfoxx add skills/agent-skills

Skills for code review and security

Codex plugin for Claude Code

OpenAI released the Codex plugin, which brings a second model's opinion into Claude Code. /codex:review checks uncommitted changes or a branch against its base. /codex:adversarial-review challenges design decisions on a topic you set, and /codex:rescue hands a task to a Codex subagent.

You need Node.js 18.18 or later, the Codex CLI and a sign-in with a ChatGPT account or an OpenAI API key. Requests count against your Codex limits. Access is extended because code and session context go to OpenAI. The review gate option on the Stop hook can loop the two agents, and then the limits run out fast. After installing, run /codex:setup in a session.

claude plugin marketplace add openai/codex-plugin-cc && claude plugin install codex@openai-codex

security-audit from Cloudflare

security-audit runs a code security audit in six phases. The agent first studies the architecture and trust boundaries. Then isolated subagents look for vulnerabilities against a coverage log, and every finding is rechecked by an agent that didn't find it. A finding is marked confirmed only after a full trace through the source.

The skill doesn't edit your code. It describes fixes in a report and a findings.json file. Builds and tests of the target run only in a sandbox with the network off. Access is extended. The authors say repeat runs find noticeably more, so run the audit two or three times before a release.

npx skillfoxx add skills/security-audit

Skills for documents, writing and marketing

Anthropic's document skills

Anthropic's skills live in the company's official repository. The docx, pdf, pptx and xlsx skills power Claude's work with Word, PDF, PowerPoint and Excel files. The same repository has mcp-builder, a template for your own skill and the Agent Skills spec. Anthropic calls the repository a reference and warns that Claude's behavior may differ from the examples. The source of the four document skills is public, but they don't have an open-source license.

Access is extended, since the skills run scripts to process files. The document-skills plugin installs only these four.

claude plugin marketplace add anthropics/skills && claude plugin install document-skills@anthropic-agent-skills

Humanizer

Humanizer removes signs of AI writing based on Wikipedia's list. The patterns fall into five groups, among them punchy closers, forced groups of three, promotional phrasing and decorative bold. The agent marks what it finds, writes a draft and checks that it didn't drop facts or add new ones. Give it a sample and it adapts to the author's voice. In a file it edits only prose and leaves code and links alone. Access is minimal.

npx skillfoxx add skills/humanizer

Marketing Skills

Marketing Skills is about 50 skills for marketing: A/B tests, landing pages, email, pricing, onboarding, SEO audits and ads. The product-marketing skill stores the product and audience description, and the others read it first. Copywriting, cro and ab-testing reference each other and work together. Access is minimal.

npx skillfoxx add skills/marketing-skills-for-ai-agents

Skills for context and memory

Planning with Files

Planning with Files moves the plan out of the context window into three project files. task_plan.md holds phases with checkboxes, findings.md holds notes and decisions, and progress.md is the session log. Hooks at session start, on every prompt and before context compaction feed the current phase back to the agent, so after /clear it picks up where it left off. A completion-gate mode keeps the agent from stopping while a phase is still open.

Access is extended, as the hooks run scripts on every step. Plan files are excluded from git by default. The SkillFoxx CLI has no automatic install for this skill yet. The command from the author's README installs the skill with its hooks.

npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g

Claude-Mem

Claude-Mem uses hooks to record what the agent did and what it decided, and stores it in SQLite. In a new session the relevant context from past ones is added automatically. You can search memory with the mem-search skill and three MCP tools: search, timeline and get_observations. Sensitive fragments are excluded with tags, and a web UI shows the memory stream.

Access is broad. Hooks record agent actions and tool output, and a worker with an HTTP API runs on your machine. By default the installer offers to sign in and keep memory in the cloud. To skip that, pass an explicit --provider flag or set CLAUDE_MEM_ONLINE_OPTIN=false.

npx claude-mem install

Claude HUD

Claude HUD puts the model, git branch, context usage and limits in the Claude Code status line, two lines by default. It takes token counts from Claude Code itself instead of estimating them. You can add lines for active tools, subagents and the todo list. HUD gives the agent nothing new, but it shows you when context is getting close to the edge. Access is minimal, the plugin only reads session data. After installing, run /claude-hud:setup.

claude plugin marketplace add jarrodwatts/claude-hud && claude plugin install claude-hud@claude-hud

Skills for UI design

I wrote a separate piece on 10 skills for UI design. Here are three to start with.

UI UX Pro Max

According to its README, UI UX Pro Max keeps a local database of 79 searchable styles, 192 palettes, 74 font pairings and 119 UX guidelines. The agent picks a style, colors and fonts by product type, writes code and checks it against the guidelines. Search runs on Python 3 with no network access, and logos and branding are in the paid version. Access is extended, the skill runs local scripts.

npx skillfoxx add skills/ui-ux-pro-max

Impeccable

Impeccable gives the agent 24 commands to critique and polish an interface, among them audit, critique, polish and typeset. Its README lists 62 deterministic rules that find template patterns in code without calling the model. Access is broad. The skill installs hooks that fire on edits and may download an engine binary to ~/.impeccable/bin on first run. Run the install command from the project root.

npx impeccable install

Taste Skill

Taste Skill reads the brief and picks an art direction before writing code. Since version 2 it's one skill, design-taste-frontend, that builds with real photos from open collections and checks the page in a real browser at three widths. It needs Node 22 or newer and Chrome, Edge or Chromium. The older style skills moved to a separate taste-skill-classics repository.

npx skillfoxx add skills/taste-skill

How do you avoid overloading Claude Code with skills?

Claude Code keeps the descriptions of all installed skills in context and opens the full SKILL.md only for a matching task. So every set costs context before its first call. Marketing Skills brings about 50 descriptions, Matt Pocock's set a few dozen. Even Caveman, which shortens the agent's answers, adds 1,000 to 1,500 input tokens per turn according to its author.

These are the rules I follow myself.

• Pick one development workflow. Superpowers, Pocock's skills and Osmani's Agent Skills describe the same cycle. Install them together and the agent gets three conflicting sets of instructions.
• Install a skill per project if only one repository needs it. A global install adds its description to every session.
• Take only the part you need from large sets. Anthropic's skills have the document-skills plugin for that, and npx skills add has the --skill flag.
• Watch your hooks. Superpowers, Planning with Files, Claude-Mem and Impeccable run scripts on session events, and each hook can add text to the context.
• Don't install the same skill two ways. The author of Pocock's skills warns that the plugin plus a copy from skills.sh will duplicate the skills.
• Once a month, open the list with /skills and remove what you haven't used. Skills installed through SkillFoxx are removed with npx skillfoxx remove skills/<name>, plugins with /plugin uninstall.

Start with three to five skills from the first section and add the next one when you hit a specific task. The rest, with access levels and install commands, are on the Claude Code page in the catalog. Plugins get their own list in the best Claude Code plugins. If you don't know what to look for, describe your project in Foxx AI matching and it will suggest skills for your files.