Claude sits on top of a massive skills library — over a thousand options when you count the community-built ones. Most of them aren’t worth your time. A lot of them are just slightly different ways to do something Claude already does fine on its own.
But a handful of them are genuinely different. One of them doubles your daily Claude usage without paying a penny more. Another does a brutal startup teardown — Y Combinator style — on any business idea in under 40 seconds. And one of them is the exact setup that Garry Tan, the CEO of Y Combinator, published on GitHub after using it to average 10,000 lines of code and 100 pull requests per week.
This post covers the 10 skills worth installing, how to actually set them up (with real commands, not vague instructions), and why the skills system works the way it does. It’s based on Vaibhav Sisinty’s May 2026 breakdown of the best Claude skills — plus deeper research into the gstack architecture and setup process.
What Claude skills actually are (and why most people misunderstand them)
A Claude skill is a folder with a SKILL.md file inside it. That file has two parts: a short YAML header at the top that tells Claude when to use the skill, and markdown instructions in the body that tell Claude how to use it.
The architecture is smarter than it sounds. Claude doesn’t load every skill into memory at once. It reads only the name and description from each skill when a session starts — about 100 tokens per skill. When you give it a task, it checks whether any skill descriptions match the situation. If one does, it loads the full skill content. If none do, they stay dormant. This is called progressive disclosure, and it means you can have 50+ skills installed without any slowdown on unrelated tasks.
There are two kinds of skills and understanding the difference saves a lot of confusion:
| Type | What it does | Example |
|---|---|---|
| Capability Uplift | Gives Claude abilities it doesn’t have natively | Creating real PDF files, browser automation, web scraping |
| Encoded Preference | Teaches Claude your team’s specific way of doing something it already knows | Your commit message format, your brand voice, your code review checklist |
Most beginners go looking for Capability Uplift skills when Encoded Preference skills are often where the real value is. The difference between Claude giving you a generic output and Claude giving you exactly what your workflow needs is almost always an Encoded Preference skill.
How to install Claude skills: 3 ways (with actual commands)
Before getting to the list, here’s how to actually install skills — because the instructions scattered across the web are often vague or outdated.
Method 1: The plugin browser inside Claude (easiest)
If you’re using Claude Code, open a session and type /plugin. This opens the plugin browser. Go to the Discover tab, find the skill you want, and press Enter. Claude will ask whether to install it for all projects (User) or just the current folder (Project). That’s it.
For official Anthropic skills, you can also install directly by name:
/plugin install frontend-design@anthropic-agent-skillsMethod 2: One-line curl install (for Agensi-hosted skills)
Agensi is the largest skills marketplace and security-scans every submission before listing it. For any skill hosted there, you can install with a single line in your terminal:
mkdir -p ~/.claude/skills && curl -sL https://www.agensi.io/api/install/SKILL_SLUG | tar xz -C ~/.claude/skills/Replace SKILL_SLUG with the skill name from the Agensi page. The command creates the folder, downloads the skill, and unpacks it in one step. Type /skills inside a Claude Code session to confirm it loaded.
Method 3: Manual install from GitHub (for custom or community skills)
For skills from GitHub repos or ones you’ve built yourself:
# Personal install — available across all your projects
cp -r my-skill-folder/ ~/.claude/skills/
# Project install — current repo only (shared via git)
cp -r my-skill-folder/ .claude/skills/
# Verify it loaded
/skillsRestart your Claude Code session after copying. To remove a skill, delete the folder. To temporarily disable one without deleting it, rename the folder with a leading underscore: _my-skill/
For Claude.ai (browser or desktop app)
Go to Settings → Customize → Skills → Upload. Zip the skill folder first — the ZIP needs to contain the folder itself at the root, not just the SKILL.md file. After uploading, Claude reads it and shows you the skill name and description. Done.
The 10 best Claude skills to install in 2026
Here they are, in order of immediate practical impact. The install command is included for each one.
1. Caveman — the one that doubles your usage for free
This is the skill Vaibhav calls out first, and for good reason. Caveman mode compresses Claude’s responses by roughly 75% while keeping full technical accuracy. Claude responds in ultra-short, punchy sentences — no padding, no filler, no lengthy explanations you didn’t ask for.
Why does this double your subscription? Because most of Claude’s token usage goes toward generating responses, not reading your prompts. Cut response length by 75% and you effectively get 4x more back-and-forth in the same usage window. You’re not getting less information — you’re cutting the parts that weren’t adding anything.
Trigger it by typing /caveman or just saying “caveman mode” or “be brief.” It supports intensity levels: lite, full (default), and ultra if you want it taken further.
/plugin install caveman@anthropic-agent-skills2. gstack — Garry Tan’s exact Claude Code setup (23 tools, MIT license)
Garry Tan is the CEO of Y Combinator. In March 2026, he open-sourced his full Claude Code setup on GitHub. The repo hit 97,000+ stars. The setup is called gstack.
gstack gives you 23 slash commands that turn Claude Code into a virtual development team. Each command embodies a specific role — CEO, Designer, Engineering Manager, QA Lead, Release Engineer, Chief Security Officer — with defined priorities, constraints, and outputs. Work moves through a fixed loop: Think → Plan → Build → Review → Test → Ship → Reflect.
The commands most people use first:
/office-hours— six forcing questions that reframe your product idea before any code gets written/plan-ceo-review— scope review that kills bad work early/review— Staff Engineer persona scans for production bugs and applies fixes/qa— opens a real Chromium browser, runs your app, finds bugs, writes regression tests/ship— syncs, tests, audits, pushes, and opens a PR in one chain/cso— OWASP Top 10 + STRIDE threat modeling in one pass
Install with one command:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git
~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setupYou need Claude Code, Git, and Bun v1.0+. Claude Pro ($20/month) is required for Claude Code itself.
One caveat worth saying plainly: the criticism of gstack is fair. Each skill is markdown with system prompt instructions — you could write similar files yourself. What you’re getting is a finished, opinionated, battle-tested set from someone shipping at a rate most people won’t match. Whether that’s worth more than your own half-finished prompts is a decision only you can make.
3. frontend-design — the official Anthropic skill with 277,000+ installs
This is the most-installed skill in the official Anthropic marketplace. It gives Claude a design system before it touches code, which produces actually distinctive UI instead of the default “Inter font, purple gradient” output that most LLMs converge on without direction.
If you’ve ever asked Claude to build a UI and felt like everything looks slightly the same — this is why, and this skill is the fix.
/plugin install frontend-design@anthropic-agent-skills4. skill-creator — the meta-skill that builds other skills
Skill Creator is a skill whose job is to build other skills. You describe the workflow you want to automate, it runs an interactive Q&A, and it generates a complete SKILL.md file with proper frontmatter and instructions.
The useful part: it splits your example prompts into a train/test set, measures how reliably the skill triggers, generates improved descriptions, and picks the best one by score. This solves the most common skill problem — descriptions that are too vague to trigger consistently.
If you use Claude Desktop or Claude Cowork, Skill Creator comes pre-installed. For Claude Code:
/plugin install skill-creator@anthropic-agent-skills5. YC teardown skill — brutal startup feedback in 40 seconds
This is one of the skills in the gstack pack, but worth calling out separately because the use case is distinct. The /office-hours command runs a Y Combinator-style forcing-question session on any business idea. Six questions that are designed to expose the weakest parts of a pitch before you’ve wasted time building the wrong thing.
It’s not gentle. That’s the point. Getting a brutally honest tear-down on a business idea in 40 seconds is more useful than a supportive brainstorm session that leaves you feeling good about a bad idea.
Installed as part of gstack (see #2 above).
6. humanizer — removes AI writing patterns from text
Based on Wikipedia’s documented “Signs of AI writing” guide, this skill scans text for the patterns that make AI-generated writing detectable — the em dash overuse, the hollow “testament to” phrases, the rule-of-three structures, the promotional language — and rewrites them into natural prose.
Useful for anything going public: blog posts, emails, LinkedIn content, pitch decks. The output isn’t just “less AI-sounding” — it’s more specific, more varied in rhythm, and more readable.
/plugin install humanizer@anthropic-agent-skills7. docx — Word documents with real formatting
A Capability Uplift skill. Without it, when you ask Claude to produce a Word document, you get markdown that you’d have to copy and format yourself. With it, Claude generates actual .docx files with proper headings, tables, page numbers, and formatting — ready to open in Microsoft Word or Google Docs.
/plugin install docx@anthropic-agent-skills8. pptx — proper PowerPoint decks, not slide-shaped markdown
Same principle as docx, applied to presentations. Without this skill, Claude can describe what a slide deck would look like. With it, Claude produces an actual .pptx file with real slides, layouts, and formatting you can open in PowerPoint or Google Slides.
/plugin install pptx@anthropic-agent-skills9. pdf — actual PDF files, not print-to-PDF workarounds
Claude can’t natively produce a properly formatted PDF. This skill gives it that ability — useful for reports, invoices, legal documents, or anything where you need a file that looks right and stays formatted correctly when shared. It handles text, tables, page breaks, headers, and footers.
/plugin install pdf@anthropic-agent-skills10. Superpowers (obra/superpowers) — community-built battle pack
A community-maintained collection of 20+ skills covering TDD workflows, systematic debugging, multi-agent orchestration, and performance analysis. One of the most starred community skill packs. Worth installing if you’re doing any serious development work with Claude Code.
/plugin marketplace add obra/superpowers-marketplaceHow to build your own skill from scratch
If none of the existing skills quite fit your workflow, building your own takes less than 30 minutes. Here’s the structure:
my-skill/
├── SKILL.md ← required — this is the whole thing for simple skills
├── scripts/ ← optional — Python/Bash Claude can run
└── references/ ← optional — docs loaded into contextThe folder name must be kebab-case. The file must be named exactly SKILL.md — case-sensitive. skill.md or Skill.md will both be silently ignored.
The SKILL.md file has two parts. First, the YAML frontmatter:
---
name: my-skill-name
description: >
What this skill does, when to use it, and what it produces.
Be specific — this is what determines whether Claude loads it.
allowed-tools: Read, Glob, Bash
---Then the markdown body — step-by-step instructions, output format requirements, and any constraints. Keep it focused and under 5,000 words. Every line is a recurring token cost when the skill loads.
The most important thing to get right is the description field. Claude reads this to decide whether to load the skill at all. A vague description — “helps with writing tasks” — means the skill almost never triggers. A specific description — “Rewrite blog post drafts to remove AI writing patterns and match conversational brand voice. Use when user asks to humanize, edit, or de-AI text.” — means it triggers reliably.
Skills vs MCP servers vs slash commands — which should you use?
These three things overlap enough that people regularly confuse them. Quick version:
| Tool | What it is | When to use it |
|---|---|---|
| Skills | SKILL.md instructions Claude follows automatically | Recurring workflows and task-specific behaviours |
| MCP servers | External tools and data sources Claude can call | When Claude needs to connect to an API, database, or external service |
| Slash commands | Explicit user-triggered commands | When you want to decide when something runs, not Claude |
The analogy that makes it click: MCP is the kitchen — the tools, the ingredients. A skill is the recipe that tells Claude how to use them. You can combine them. A skill handles the workflow logic; MCP handles the data access. But for most use cases, a skill alone is enough to start.
Where to find skills
- Official Anthropic skills — github.com/anthropics/skills — Apache 2.0 license, security-reviewed
- Agensi marketplace — agensi.io — largest community marketplace, security-scanned before listing
- Awesome Claude Skills — github.com/travisvn/awesome-claude-skills — curated community list
- SkillsMP — skillsmp.com — broader agent skills marketplace
One security note worth taking seriously: skills can execute code in Claude’s environment. Only install from sources you trust. For anything from a less-known source, open the SKILL.md and any scripts in the scripts/ folder and read them before enabling. Pay attention to Bash commands and network requests.
Frequently asked questions
Are Claude skills free?
Yes. The SKILL.md format is open and free. Official Anthropic skills on GitHub are Apache 2.0 licensed. Community skills are free to use. A small number of premium skills on third-party marketplaces have a price, but the vast majority of the ecosystem is free — including all 10 skills in this list.
Do skills work on Claude.ai (browser) and Claude Code?
Both, but the installation method differs. Claude.ai: Settings → Customize → Skills → Upload (zip the folder first). Claude Code: copy the folder to ~/.claude/skills/ for personal installs or .claude/skills/ in your repo for project-scoped installs. The SKILL.md format works on both.
Can I have multiple skills active at once?
Yes. Claude can load multiple skills simultaneously and they’re designed to work alongside each other. The progressive disclosure architecture keeps unused skills out of context, so having many installed doesn’t slow things down on unrelated tasks.
What’s the most common reason a skill doesn’t trigger?
The description in the YAML frontmatter. If it’s too vague, Claude doesn’t recognize the task as a match for the skill. Most skill failures aren’t instruction failures — they’re description failures. Tighten the description with specific trigger phrases and it usually fixes itself.
Does gstack require anything special beyond Claude Pro?
You need Claude Code (requires Claude Pro at $20/month), Git, and Bun v1.0+. Node.js if you’re on Windows. The setup script handles the rest.
The honest summary
Skills are the most underused part of Claude. Most people who’ve been using Claude for months haven’t installed a single one. They keep pasting the same instructions into chat, getting slightly inconsistent outputs, and wondering why.
The caveman skill alone is worth installing today if you’re on a usage-limited plan. The gstack pack is worth a look if you’re doing any serious development work. And if you find yourself writing the same setup instructions at the start of every Claude session, that’s a skill waiting to be written.
None of this requires a bigger subscription. The skills themselves are free. The setup for each one takes under 5 minutes. And the difference between Claude giving you generic output and Claude giving you exactly what your workflow needs is usually a single well-written SKILL.md file.