AI: Skill

AI skills are reusable “recipes” that an agent can load when needed: instructions for a task, optionally accompanied by scripts, references, and assets. The Agent Skills format packages these in a folder with a SKILL.md file so they can be reused across skills-compatible agents.

For Claude Code’s installation paths, commands, and plugin setup, see Skills.

Creating skills#

The common format is YAML frontmatter followed by Markdown instructions. The specification requires two fields:

• name: the skill’s identifier.
• description: what the skill does and when to use it.

A skill can consist entirely of instructions. Supporting files are optional:

math/
├── SKILL.md
├── scripts/
│   └── verify.py
├── references/
└── assets/

Example: a minimal math skill#

---
name: math
description: >
  Verify mathematical derivations with SymPy.
  Use when checking proofs, calculations, or symbolic mathematics.
---

When asked to verify a mathematical derivation:

1. Use the bundled scripts/verify.py to check the expressions.
2. Explain the reasoning alongside the result.
3. State what was checked and any assumptions.

The script supplies the calculation; the instructions describe when and how to use it. A skill with no executable code can instead encode a checklist, a writing convention, or another repeatable procedure.

How skills load#

The format uses progressive disclosure:

1. Discovery: the agent sees the available skills’ names and descriptions.
2. Activation: it reads a relevant skill’s full instructions.
3. Execution: it loads supporting files or runs bundled code as needed.

This keeps detailed procedures out of the initial context until they are useful. Installation, direct invocation, and extra frontmatter options depend on the host; Skills covers Claude Code’s implementation.

Skill collections and examples#

• anthropics/skills - Official repository with example skills from Anthropic
• richard-kim-79/archora-skills—research-pipeline skills (hypothesis, validation, IMRaD draft, figure, stats, counterargument, peer-review simulation, synthesis, wiki-sync); companion to the Archora platform
• badlogic/pi-skills—skills for the pi coding agent, also compatible with Claude Code and Codex CLI (brave-search, browser-tools, Google Calendar/Drive/Gmail, Whisper transcription, VS Code, YouTube transcripts)
• mattpocock/skills—Matt Pocock’s skills, including caveman, grill-me, and grill-with-docs.
• stephenturner/skill-nih-biosketch—Stephen Turner’s standalone skill that drafts NIH biosketch (Common Form) content for SciENcv from a CV and grant documents. See skill nih biosketch.
• skojaku/snakemake-skill—Sadamori Kojaku’s skill encoding his Snakemake style conventions (path constants, parameter combinations, named inputs/outputs, dual-mode scripts, rule inheritance).
• skojaku/research-project-skill—Sadamori Kojaku’s skill for managing a research project’s lifecycle: scaffold experiments, track them as GitHub issues, mark results paper-ready.

Claude Code-focused collections, plugin suites, and harness configurations are listed on Skills. Check each collection’s setup instructions for its supported hosts and dependencies.

Workflow examples#

• Show Us Your Agent Skills—demonstrations of practical agent skills and workflows.