The Best Claude Code Skills in 2026

find-skills turns the agent into a skill scout. Ask "how do I do X" or "is there a skill for X", and it checks the skills.sh leaderboard first. Then it searches with npx skills find <query>, from Vercel's Skills CLI.

The useful part is its quality bar. Before recommending anything, the agent checks three things. Installs: prefer 1K+, be careful under 100. Source: official publishers first. Stars: be skeptical of repositories under 100. It then shows you the skill, its installs, the command and a skills.sh link.

Know what it does next. If you say yes, it installs with npx skills add <owner/repo@skill> -g -y, which is global and skips the confirmation prompt. If you'd rather review a skill before it lands, ask the agent to stop at the recommendation.

grill-me is "a relentless interview to sharpen a plan or design". You call it with /grill-me, since it sets disable-model-invocation: true and Claude never runs it on its own.

The folder is a one-line shortcut. Its only instruction is to call another skill, grilling, from the same repository. So install both, or the command has nothing to run: skillpack install github:mattpocock/skills/skills/productivity/grilling --scope project.

grilling is where the method lives. The agent maps your plan as a design tree. Each round, it asks every question that can be settled now, numbered, with its recommended answer. Questions are worded so that "yes" accepts the recommendation. Facts it can look up itself, it sends a sub-agent to find. It doesn't act until you agree the tree is fully explored.

agent-browser lets the agent open pages, fill forms, click, take screenshots and pull data from websites. It drives Chrome or Chromium over the DevTools protocol, through a Rust CLI with no Playwright or Puppeteer dependency. Pages come back as accessibility-tree snapshots with short element refs like @e3.

The skill itself is a stub. It tells the agent to install the CLI (npm i -g agent-browser && agent-browser install) and then load the real instructions from it with agent-browser skills get core. Vercel's reasoning is that the CLI serves guides that always match the installed version. Extra guides cover Electron apps, Slack, exploratory QA and cloud browsers.

Two things to know. Its frontmatter pre-approves agent-browser commands through allowed-tools. Its description also tells the agent to prefer it over any built-in browser tool, so it will take over browsing tasks once installed.

HyperFrames is HeyGen's open-source framework for rendering video from HTML, and this skill is its front door. It calls itself the "mandatory entry point" for any request to make, edit, animate or render a video. Its sister skill hyperframes-cli has slightly more installs (1.18M), but it's the reference for the CLI, so we list the entry point.

For a new video, it runs a short intent interview and writes a BRIEF.md. It then routes to one of ten workflows, from product launch videos and explainers to captions, PR-to-video and Remotion ports.

It depends on the HyperFrames CLI, which installs each workflow with npx hyperframes skills update <workflow>. Your Skillpack lockfile doesn't track those workflow files. The folder also ships Node scripts, which the Skillpack CLI lists before it copies anything.

frontend-design is Anthropic's answer to generic, templated UI. It casts Claude as the design lead of a studio whose client "has already rejected proposals that felt cliché or templated". It asks for deliberate choices about palette, typography and layout.

The rules are specific. The agent identifies the product and audience before designing, keeps line lengths under 80 characters, and avoids the usual tells of a generated page. Those include accenting one word in a headline, all-caps labels, and numbered markers on content that isn't a sequence. It then plans, builds and critiques its own work against the brief.

It's a direction skill, not a component library or a checklist. It won't pick your design system for you.

caveman changes how the agent talks, not what it does. Its description sums it up: "answer first, fluff gone, every technical fact kept". Turn it on with /caveman or "be brief", and it stays on for the session until you say "stop caveman" or "normal mode".

The name oversells the gimmick. The rules ban greetings, recaps, hedging and filler words, and cap sentences at 20 words. They also forbid dropping "not", "never" or "only", and keep code, commands, paths and errors verbatim. It switches back to full sentences for security warnings, irreversible actions and questions to you.

Anything written outside the chat stays in normal prose: code, comments, commits, docs and PRs. The /caveman ultra and /caveman wenyan modes point to other skills in the same repository, which this folder doesn't include.

brainstorming is the first step of Superpowers, Jesse Vincent's development methodology. Its description says the agent "MUST use this before any creative work". Its premise is that the agent's idea of what you want "is usually thinner than it feels".

The agent reads the code and anything you pointed at before its first question. It then asks about the one thing your request leaves open and lays out two or three approaches for you to pick from. Next comes a design doc, short for a small change and full for a project. Nothing gets built until you approve that doc, which it saves under docs/superpowers/specs/.

It's part of a chain. Once the design is approved, it calls the writing-plans skill from the same repository, so install that too or use the whole Superpowers set. The folder also ships a small local server for an optional visual companion, which the Skillpack CLI lists as executable files.

seo-audit is the most-installed skill from Marketing Skills, Corey Haines' collection for marketers. It triggers on phrases like "why am I not ranking", "my traffic dropped" or "stuck on page 2".

It audits in a fixed priority order: crawlability and indexation, technical foundations, on-page optimization, content quality, then authority. Every finding comes back with its impact, evidence, fix and priority, ending in an action plan. It reads a .agents/product-marketing.md file first if you have one, so it doesn't ask what it already knows.

It's honest about its limits. It warns that curl and web fetches can miss JSON-LD added by JavaScript, and sends you to the Rich Results Test for schema. It also treats fetched pages as untrusted data, a sensible guard against prompt injection.

playwright-cli is Microsoft's skill for the Playwright CLI. The agent opens a browser, navigates, clicks and fills elements by ref, and works with tabs, storage, network and DevTools from short commands.

Microsoft's README explains why it exists next to Playwright MCP. CLI calls are more token-efficient, since the agent doesn't load large tool schemas or full accessibility trees. MCP stays the better fit for long, stateful sessions like self-healing tests.

You need Node.js 18 or newer and the CLI itself (npm install -g @playwright/cli@latest). Its frontmatter pre-approves playwright-cli and npx playwright commands through allowed-tools.