A Claude skill is a folder. Inside it, a file called SKILL.md with a YAML frontmatter block (name, description) and a body of markdown instructions. Claude reads the description, decides whether the user's request matches, and pulls the body into context if it does. That's the whole format. No SDK, no plugin runtime, no compile step.
The format reads simple on paper. You skim the spec and think "okay, prompt-engineering with a folder." Then you open Anthropic's pdf skill and find a 200-line reference doc, three Python helper scripts, a forms sub-guide, and a description that triggers on the literal substring ".pdf". The shape of a good skill is hard to grasp until you see a dozen of them next to each other.
So that's what's here. Fifteen skills published in public GitHub repos as of May 2026, the SKILL.md description pasted from the actual file, the author, the link, and an honest take on whether the skill earns its context cost or is a demo someone forgot to delete.
Document handling: the four flagship skills
The most-installed skills in the wild are Anthropic's four document-format skills. They live at github.com/anthropics/skills, they're source-available rather than Apache-licensed, and they're the reason "Claude can edit a Word doc" stopped being a feature request and became a one-line install.
Start with pdf. Its description reads: "Use this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable." It also says, plainly, "If the user mentions a .pdf file or asks to produce one, use this skill." That last line is the trick. A file-extension substring is about as cheap and reliable a trigger as you can write. Install it. The skill pays for its context cost the first time someone hands you a 60-page scanned PDF and you don't have to remember which Python library handles OCR.
Same idea for Word files: docx. The description is longer, sharper about when to trigger ("any mention of 'Word doc', 'word document', '.docx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads") and explicit about when not to ("Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation"). The body walks Claude through unpacking the .docx XML, editing it, and repacking. It's the skill that finally makes "draft me a memo" produce a real Word file instead of a markdown blob.
"Every Excel model MUST be delivered with ZERO formula errors." That's the first line of xlsx's body. If you've ever shipped a financial model with a #REF! in column F, you know why. The frontmatter also spends a paragraph defining when a spreadsheet is the primary deliverable, because the failure mode is Claude reaching for it during any tabular task. Worth installing if you produce spreadsheets weekly. Skip if you mostly read them.
pptx is the presentation one and took the longest to feel good in practice. Decks are visually load-bearing in ways PDFs aren't, and the first generation of "Claude makes a slide deck" produced things that looked like a 2004 Keynote default template. The current skill routes through three sub-guides depending on whether the task is reading, editing a template, or generating from scratch with pptxgenjs. Triggers on "deck," "slides," "presentation," or any .pptx filename. Much better than before.
Authoring and infrastructure skills
Two skills in the official repo are about building more skills. They're recursive but useful.
Every Claude Code user hits the same wall after their first week: "how would I even test whether my skill is working?" skill-creator answers it. The frontmatter triggers on drafting a skill, editing an existing one, running evals, benchmarking with variance analysis, or tuning a description for better triggering accuracy. The body is a six-step loop: draft, write test prompts, run them, evaluate qualitatively and quantitatively, rewrite, repeat. It ships an eval-viewer script that renders side-by-side runs. Anthropic's answer to "is my skill any good" is that you write twenty prompts and grade the outputs. If you're going to author skills, install it.
mcp-builder does the same job for MCP servers. Description: "Guide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK)." The body lays out a four-phase workflow, narrower than skill-creator because the MCP spec is tighter. The guidance on tool-name length, parameter shape, and error-response contracts is the kind of thing you'd otherwise learn by shipping three bad MCP servers first.
Code-quality skills (community-built)
The two skills I install on every fresh machine come from Jesse Vincent's Superpowers framework, a much larger collection of agentic skills that's also distributed as a plugin.
"Use when implementing any feature or bugfix, before writing implementation code." That's the entire description for test-driven-development. The body opens: "Write the test first. Watch it fail. Write minimal code to pass." Then: "If you didn't watch the test fail, you don't know if it tests the right thing." With it installed, Claude stops jumping straight to implementation and asks what the failure case looks like. The reason this works as a skill rather than ambient system-prompt rule is that the description fires on a specific moment ("any feature or bugfix"). The five-token cost pays back the first time Claude writes a regression test before patching the bug.
Its sibling systematic-debugging opens with "Random fixes waste time and create new bugs" and triggers on "any bug, test failure, or unexpected behavior, before proposing fixes." It enforces a root-cause-first workflow. Without it, Claude will happily wrap a try/except around a KeyError and call the bug fixed. With it, Claude asks why the key is missing in the first place.
Style and brand
The most underrated skill in Anthropic's repo is frontend-design. Description: "Create distinctive, production-grade frontend interfaces with high design quality... Generates creative, polished code and UI design that avoids generic AI aesthetics." The body opens with a list of aesthetic directions ("brutally minimal, maximalist chaos, retro-futuristic, organic/natural, luxury/refined, playful/toy-like, editorial/magazine, brutalist/raw") and tells Claude to pick one and commit. That paragraph is the entire fix for what frontend code from Claude looked like in 2024: a hundred variations of a centered card with a purple gradient and a rounded button. Install it even if you don't think you need it. The default output without it is the AI-design tell.
Domain-specific: legal and finance
Anthropic shipped a separate repo, knowledge-work-plugins, aimed at non-engineering workflows. It's organized by department: legal, finance, sales, marketing, HR, ops. Most of the skills carry a disclaimer that they assist with workflows but don't replace a qualified professional, which is correct and worth pasting into the top of any domain skill you build yourself.
A use case from the legal set: reviewing vendor agreements against your in-house playbook. review-contract description: "Review a contract against your organization's negotiation playbook, flag deviations, generate redlines, provide business impact analysis. Use when reviewing vendor or customer agreements, when you need clause-by-clause analysis against standard positions, or when preparing a negotiation strategy with prioritized redlines and fallback positions." What makes it work is the playbook hookup. The plugin ships a CLAUDE.md that gets populated through a guided interview where the user supplies signed contracts and internal standard positions. The skill then reviews against your playbook, not a generic one. That's the gap between "Claude reviews a contract" and "Claude reviews a contract the way your in-house counsel would." If you scan this list and think "I could ship something like this for my niche," Knack was built for exactly that, and the no-code build walkthrough is the next step.
On the finance side, variance-analysis is the one I keep coming back to. Description: "Decompose financial variances into drivers with narrative explanations and waterfall analysis. Use when analyzing budget vs. actual, period-over-period changes, revenue or expense variances, or preparing variance commentary for leadership." The skill knows the difference between a price-volume-mix decomposition and a budget-vs-forecast walk, which is domain texture you'd never expect Claude to bring unprompted. If you write monthly variance commentary, install it.
Stripe's integration skills
Two skills from Stripe's agent-toolkit repo show what a vendor-authored skill should look like.
stripe-best-practices is the long one. Its description names the actual decisions: "API selection (Checkout Sessions vs PaymentIntents), Connect platform setup (Accounts v2, controller properties), billing/subscriptions, Treasury financial accounts, integration surfaces (Checkout, Payment Element), migrating from deprecated Stripe APIs, and security best practices (API key management, restricted keys, webhooks, OAuth)." That description is itself the most useful part. It tells the model exactly which decision points Stripe has an opinion about. The companion upgrade-stripe is a one-job tool: bump your Stripe SDK version safely, with the current API version pinned in the first line of the body. The vendor knows which mistakes their integrators make. The skill encodes that knowledge.
Two honourable mentions
"Why use many token when few token do trick." That's the tagline of Julius Brussee's caveman repo, and the skill does what it says. Description: "Ultra-compressed communication mode. Cuts token usage ~75% by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra." The wenyan modes compress through classical Chinese, which sounds like a joke and works. Install it if you burn through Claude's context window on long sessions. Skip it if you're shipping output to a human reader who isn't in on the bit.
Vercel Labs ships the meta-skill: find-skills. Description: "Helps users discover and install agent skills when they ask questions like 'how do I do X', 'find a skill for X', 'is there a skill that can...', or express interest in extending capabilities." It hooks into npx skills and turns "I wish Claude could do X" into a search-and-install loop. The cost of looking for a skill that doesn't exist is much lower than the cost of writing one that already did.
What the good ones share
Look at the descriptions side by side. The good skills name a specific moment: "any time a .pptx file is involved," "before writing implementation code," "when reviewing vendor or customer agreements." The bad ones, the kind that fill the marketplace's long tail, say things like "helps with productivity" or "assists in development tasks." A skill description is a routing decision, and routing decisions need concrete tokens to match against. "When the user mentions a .pdf file" is a token match. "When the user is being productive" is nothing.
The good skills also know how to say no. xlsx devotes a paragraph to when not to use itself. The claude-api skill tells Claude to skip itself entirely if the project file imports the OpenAI SDK. The bad skills want to be invoked for everything. They sprawl. They list ten categories in the description and produce mediocre output in all of them. The average skill in the wild is too broad. The ones that work do one specific job extremely well, and the description tells the model exactly when that job is happening.
If you're about to write your first skill, install five of the ones above and read their SKILL.md files in full. Most of the engineering is in the description. The body is the easy part.