Claude Skills landed in October 2025 and the ecosystem went from zero to genuinely useful in about a week. Skills are just folders — a SKILL.md with instructions, optional scripts, optional reference docs — that Claude loads on demand when the task matches. No fine-tuning, no MCP server to host, no API gymnastics. Drop a folder in, Claude reads it when relevant.
Most of the good ones are free on GitHub. Here's what's actually worth installing.
How Skills work (the 30-second version)
A Skill is a directory. At minimum it contains a SKILL.md with YAML frontmatter — a name and a description. The description is the important bit: Claude uses it to decide whether to load the skill for a given task. Inside the folder you can ship Python scripts, templates, reference markdown, whatever. Claude reads the metadata always, the body when triggered, and the scripts when it needs to run them.
This is "progressive disclosure" — Anthropic's term, and it's the right design. Your context window doesn't get burned on skills you aren't using.
You install them by dropping the folder into ~/.claude/skills/ (Claude Code) or uploading via the Skills UI (claude.ai, Pro/Max/Team/Enterprise).
Anthropic's official skills
Start here: anthropics/skills. This is the reference repo and it's better than most third-party stuff.
The document skills are the headline act:
- pdf — fill forms, extract text, merge, split. Uses
pypdfunder the hood. - docx — generate real Word documents with tracked changes, comments, styles. Not "export HTML and rename it".
- xlsx — write actual Excel files with formulas, formatting, multiple sheets.
- pptx — generate PowerPoint decks. Surprisingly good at layout.
If your work touches Office formats at all, install these four and stop writing brittle export scripts.
Also in the repo: artifacts-builder (for building richer Claude artifacts), mcp-builder (scaffolds MCP servers), skill-creator (a skill for writing skills — meta but useful), and webapp-testing for Playwright-driven UI checks.
obra/superpowers
obra/superpowers by Jesse Vincent is the most ambitious community pack. It's not one skill — it's dozens, organised around a methodology. Brainstorming, planning, TDD workflows, debugging heuristics, root-cause analysis, condition-based waiting instead of sleep(), defensive programming patterns.
The opinionated bit: it pushes Claude toward a specific way of working — write the test first, verify the hypothesis before changing code, never commit without running the suite. If you already work that way, it amplifies you. If you don't, it'll frustrate you for a day and then probably convert you.
Install via their plugin marketplace command or clone directly into ~/.claude/skills/.
wshobson/agents and wshobson/commands
Not strictly Skills, but worth mentioning in the same breath: wshobson/agents is a large collection of subagent definitions, and wshobson/commands ships slash commands for Claude Code. Both convert cleanly to the Skills format if you want, and many people are doing exactly that.
The agents repo in particular has solid specialist prompts — security review, performance analysis, refactoring passes — that work well loaded as Skills.
Practical picks by use case
A few specific ones worth grabbing:
Writing and editing
The document-skills from Anthropic's repo cover most of it. For prose specifically, the superpowers writing-related skills (clear-thought, condition-based-waiting applied to research) hold up. There's also a growing set of "editor" style skills floating around — search GitHub for claude-skill topic.
Code review and PR work
Pull the relevant agents from wshobson/agents and convert. The code-reviewer, security-auditor, and test-writer ones are the highest signal. Pair with the webapp-testing skill from Anthropic for end-to-end coverage.
Data and spreadsheets
xlsx from Anthropic plus a small custom skill that knows your schema beats any generic "data analyst" prompt. The pattern: ship a reference.md with your column definitions and a scripts/validate.py that checks output. Took me an hour, saves me hours weekly.
MCP development
mcp-builder from Anthropic. If you're writing MCP servers, this scaffolds them correctly the first time instead of after three rounds of "actually the transport should be...".
What makes a skill actually good
After installing a lot of these, the pattern for the ones that earn their keep:
- Tight description. The frontmatter description is the routing layer. Vague descriptions mean the skill never triggers, or triggers wrongly. "Helps with documents" is useless. "Fills PDF forms, extracts text from PDFs, merges and splits PDF files" is correct.
- Scripts over prose. If the task is deterministic — parse this, validate that, format the other — ship a Python script. Don't make Claude reimplement it from instructions every time.
- Reference files, not stuffed SKILL.md. Put schemas, examples, edge cases in separate files Claude can pull in when needed. Keep the main file scannable.
- One job. A skill that does "everything for marketing" loads constantly and helps with nothing. Skills should be narrow.
The Anthropic skills follow this. superpowers mostly does. A lot of the long tail on GitHub doesn't — you'll see 4000-word SKILL.md files that try to teach Claude general programming. Skip those.
Where to find more
- GitHub topic:
claude-skillsandclaude-skill - The official skills repo discussion tab
awesome-claude-skillslists are starting to appear — quality varies
The takeaway
Install Anthropic's document skills today. Add superpowers if you want an opinionated workflow upgrade. Then write your own — Skills are small enough that a custom one for your codebase, your schema, your style guide will outperform anything generic. The whole point of the format is that it's cheap to make. Use that.