anthropics/skills
is a public collection of Agent Skills: portable, task-focused instruction packages that help coding agents handle repeatable workflows such as creating documents, PDFs, spreadsheets, and slides.
The repository’s key idea is refreshingly simple: put domain knowledge inside a SKILL.md
file, keep related scripts or reference files beside it, and let an agent load the right workflow when the task matches. Instead of pasting a giant prompt every time you need a report or spreadsheet, you version a reusable skill alongside your project.
Start by inspecting the available skills locally:
git clone --depth 1 https://github.com/anthropics/skills.git
cd skills
find . -name SKILL.md -maxdepth 3
A typical skill directory follows this shape:
my-skill/
├── SKILL.md
├── scripts/
│ └── validate_output.py
└── references/
└── style-guide.md
For a project-level workflow, create a skill in the location supported by your agent or IDE. The important contract is the SKILL.md
file: it should define when the skill applies, the expected output, validation steps, and any tooling constraints.
---
name: api-review
description: "Review API changes for compatibility, error handling, and documentation gaps."
---
1. Read the changed API routes and schema definitions.
2. Identify breaking request or response changes.
3. Produce a markdown checklist with file references.
4. Do not modify source files unless explicitly requested.
This structure is especially useful in Cursor-style IDE workflows: keep team conventions in version control, make them discoverable, and avoid relying on every developer’s personal prompt history.
The repository’s +277 stars today makes sense because Skills address a real agent-engineering problem: repeatability. A good skill combines instructions, references, and executable checks into a reviewable unit.
Before adopting Skills broadly, watch for:
SKILL.md
files can encode obsolete APIs or team conventions. Review them like production code.Treat Skills as lightweight automation playbooks. Small, focused, versioned workflows usually outperform one massive “do everything” agent prompt.