A Tour of Anthropic's 17 Official Skills — What the anthropics/skills Repo Can Do

Anthropic publishes Claude’s Agent Skills officially in a GitHub repository called anthropics/skills. These are reusable packages that teach Claude specific procedures and domain expertise after the fact. This article tours the 17 skills in the skills/ directory as of commit 9d2f1ae, reading each SKILL.md to explain what it does.

What is an Agent Skill?

An Agent Skill is, in a phrase, “a folder holding an instruction manual and a set of tools.” At its center sits a single Markdown file, SKILL.md, whose top carries YAML frontmatter like this:

---
name: pdf
description: Use this skill whenever the user wants to do anything with PDF files...
---

name is the skill’s identifier, and description is the trigger text that tells Claude “when should I use this skill.” This is the crux of Agent Skills: Claude normally sees only the description, and it reads the body of SKILL.md (along with the scripts, templates, and extra docs sitting in the same folder) only once a user’s request matches that description. This “don’t load until needed” mechanism is called progressive disclosure, a design that keeps a large roster of skills on standby while conserving context.

Below I group the 17 skills into five categories by their character. Here is the full picture first.

Skills for manipulating office files

Four skills for reading, generating, and editing Microsoft Office and PDF files. Each one switches between libraries such as python-docx and pypdf, or the approach of unpacking a file as a ZIP and editing its XML directly, depending on the task.

docx

A skill for Word documents (.docx). Recognizing that a .docx is really a ZIP archive of XML files, it switches between pandoc for reading, docx-js for creating from scratch, and “unpack, rewrite the XML, repack” for editing existing files. It can produce polished documents with formatting such as tables of contents, headings, page numbers, and letterheads, and it also handles tracked changes and comments, inserting and replacing images, and find-and-replace within a document.

pdf

A skill that handles anything you might do with PDF files. Using Python libraries led by pypdf plus command-line tools, it covers extracting text and tables, merging and splitting PDFs, rotating pages, adding watermarks, creating new PDFs, filling forms, encrypting and decrypting, extracting images, and OCR to make scanned PDFs searchable. Details for form filling live in a separate FORMS.md, and advanced features in REFERENCE.md.

pptx

A skill for PowerPoint presentations (.pptx). It fires whenever a .pptx is involved as input or output, and supports creating slide decks, extracting text from existing files, and editing from templates. It uses markitdown for reading and pptxgenjs for building from scratch, and for template editing it follows the “unpack, manipulate slides, rewrite content, repack” flow. A distinctive touch is generating slide thumbnails with thumbnail.py to grasp the layout before starting work.

xlsx

A skill for tasks whose primary input or output is a spreadsheet (.xlsx, .xlsm, .csv, .tsv). It adds columns, computes formulas, applies formatting, builds charts, and cleans up messy tabular data. For financial models it follows industry-standard color conventions (blue text for hardcoded inputs, black for formulas, green for links to other sheets in the same workbook, red for links to external files), and it explicitly requires delivering with zero formula errors such as #REF! and #DIV/0!. When updating existing templates, it stresses matching that template’s format, style, and conventions exactly.

Skills for generating design and art

A family of skills for producing visual deliverables. Some of them share a philosophy of articulating a “design philosophy” or “algorithmic philosophy” in words first, then expressing it, rather than just applying a template.

algorithmic-art

A skill for creating algorithmic (generative) art with p5.js. It generates flow fields, particle systems, noise fields, and dynamical systems while preserving reproducibility through seeded randomness. What stands out is its two-step approach: first it writes an “algorithmic philosophy” as a manifesto in a .md file, then the same Claude expresses that philosophy as p5.js code. The outputs are three files: the .md holding the philosophy, an .html interactive viewer, and the .js generative algorithm. It explicitly directs creating originals rather than imitating existing artists to avoid copyright violations.

canvas-design

A skill for creating posters, art, and static design pieces as .png or .pdf. Like algorithmic-art, it takes the two-step stance of “articulate a design philosophy, then express it visually,” but here the focus is on static visual communication through form, space, color, and composition. It starts by naming an art movement, like “Brutalist Joy” or “Chromatic Silence,” and treats text as a minimal visual accent.

frontend-design

A skill that steers you toward distinctive visual design (not something that reads as templated) when building new UI or reshaping an existing one. It acts as the design lead at a small studio, pushing you to make deliberate, opinionated choices about palette, typography, and layout. Interestingly, it names three looks that AI-generated design currently clusters around (a cream background with a serif display and a terracotta accent, a near-black background with a single bright accent, a broadsheet layout with hairline rules) and warns against defaulting to them unless the brief calls for it.

brand-guidelines

A skill for applying Anthropic’s official brand colors and typography to various deliverables. It includes the main colors (Dark #141413, Light #faf9f5, and so on), the accent colors (Orange #d97757, Blue #6a9bcc, Green #788c5d), and font choices like Poppins for headings and Lora for body text. Use it when you want an artifact to carry Anthropic’s look and feel.

theme-factory

A toolkit for applying a consistent theme to artifacts. You can apply a pairing of colors and fonts to slide decks, documents, HTML landing pages, and more. It ships 10 preset themes (including “Ocean Depths,” “Sunset Boulevard,” “Forest Canopy,” “Modern Minimalist,” and “Golden Hour”) that you show the user via theme-showcase.pdf to pick from, or you can generate a new theme on the fly. Where brand-guidelines specializes in Anthropic’s own brand, this one handles generic theming.

slack-gif-creator

A skill for creating animated GIFs optimized for Slack. It holds Slack’s dimension requirements as knowledge (128x128 for emoji, 480x480 for messages), along with an FPS range of 10–30, a color count of 48–128, and a duration under three seconds for emoji GIFs, and it generates frames with a Pillow-based GIFBuilder utility. It also includes validation tools for keeping file size down.

Skills for building and testing web artifacts

Two skills for working with claude.ai artifacts and local web apps. One handles the “build” side, the other the “test” side.

web-artifacts-builder

A set of tools for building complex claude.ai HTML artifacts, the kind involving state management, routing, or shadcn/ui components. The stack is React 18, TypeScript, Vite, Parcel (for bundling), Tailwind CSS, and shadcn/ui: you stand up a project with an init script, edit the code, and finally bundle everything into a single HTML file. It is meant for elaborate multi-component setups, not simple single-file HTML/JSX artifacts. It explicitly calls out avoiding “AI slop”: excessive centered layouts, purple gradients, uniform rounded corners, and the Inter font.

webapp-testing

A toolkit for driving and testing local web apps with Playwright. It supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs. It carries a decision tree that branches on whether the target is static HTML or a dynamic app and whether the server is already running, and it manages server lifecycle with with_server.py. The helper scripts are meant to be called as black boxes rather than read into context, a deliberate move to avoid burning through the context window.

Skills for writing and communications

Two skills for producing documents and internal communications along a defined structure.

doc-coauthoring

A skill that guides collaborative document authoring through a structured workflow. It triggers when writing PRDs, design docs, decision docs, RFCs, or technical specs, with Claude acting as an active guide through three stages. First, Context Gathering closes the gap between what the user knows and what Claude knows; next, Refinement & Structure iteratively builds each section; and finally, Reader Testing has a fresh Claude with no context read the doc, so the author can catch blind spots before anyone else reads it.

internal-comms

A skill for writing internal communications in the formats a company likes to use. It handles 3P updates (Progress, Plans, Problems), company newsletters, FAQs, status reports, leadership updates, project updates, and incident reports. It identifies the communication type from the request, loads the matching guideline under the examples/ directory (such as 3p-updates.md, company-newsletter.md, or faq-answers.md), and follows that file’s instructions for formatting, tone, and content gathering.

Skills for development and extension

Three skills that help you build applications with Claude, or extend Claude itself.

claude-api

A reference skill for building LLM-powered applications with the Claude API and the Anthropic SDK. It covers model IDs, pricing, parameters, streaming, tool use, MCP, agents, prompt caching, and token counting, detecting the project’s language and pointing to language-specific docs. Notably, it is vigilant about “API drift”: it tabulates places where a training-time prior has gone stale (for example, extended thinking moving from budget_tokens to thinking: {type: "adaptive"}) and directs you to verify against the skill’s own docs before writing code. It also guards against mixing Anthropic SDK calls into another provider’s code, such as OpenAI’s.

mcp-builder

A guide for building high-quality MCP (Model Context Protocol) servers that connect external services to LLMs. It supports both Python’s FastMCP and the Node/TypeScript MCP SDK. It carries a four-phase workflow that begins with deep research and planning, and it teaches design principles measured by “how well the server lets an LLM accomplish real-world tasks”: the tradeoff between comprehensive API coverage and workflow-specific tools, discoverability through consistent prefixes (naming like github_create_issue), and actionable error messages that guide the agent toward a solution.

skill-creator

A meta-skill for creating new skills, improving existing ones, and measuring skill performance. It runs a loop of “decide what the skill does, write a draft, run test prompts, evaluate the results both qualitatively and quantitatively, and rewrite based on feedback.” It can figure out where you are in this loop and jump in mid-stream, visualize evaluation results with eval-viewer/generate_review.py, and run a script that optimizes a skill’s description to improve trigger accuracy. It is, in short, the tooling that produces the very skills covered in this article.

Conclusion

We toured the 17 skills in the anthropics/skills repository across five categories.

• Office file manipulation: docx, pdf, pptx, and xlsx read, write, and edit Word, PDF, PowerPoint, and Excel.
• Design and art generation: algorithmic-art, canvas-design, frontend-design, brand-guidelines, theme-factory, and slack-gif-creator span code-generated art, UI design, brand application, and GIF creation.
• Web artifacts: web-artifacts-builder handles the build side and webapp-testing the test side.
• Writing and communications: doc-coauthoring and internal-comms produce writing along a structured workflow and a company’s own formats.
• Development and extension: claude-api, mcp-builder, and skill-creator help you build with Claude and extend Claude itself.

Because all of these are open source, the way each SKILL.md is written is itself an excellent textbook for designing your own skills. In particular, the use of progressive disclosure (keeping the body short and pushing detail into separate files) and the discipline of carefully spelling out trigger conditions in the description are lessons you can carry straight into your own skill work. A good place to start is to run one skill close to your work locally and observe how SKILL.md changes Claude’s behavior.

That’s a field note from someone who folds Claude Code and Agent Skills into daily work, on the overall shape of Anthropic’s 17 official skills.