On December 18, 2025, Anthropic published Agent Skills as an open standard, with a specification and documentation at agentskills.io. Skills package instructions, scripts and reference files into a folder that an agent loads only when a task needs it. Anthropic had introduced the format for Claude two months earlier. Opening it up turned a Claude feature into a portable way to teach any compatible agent how your team does things, and the format has since been adopted by dozens of agent products.

Key Facts#

  • Open standard date: December 18, 2025, announced as an update to Anthropic's original Agent Skills post, with the standard published at agentskills.io.
  • Original launch: Agent Skills debuted on October 16, 2025 for the Claude apps, Claude Code, the Claude Developer Platform (with a new /v1/skills API endpoint) and the Claude Agent SDK.
  • Format: A skill is a directory with a required SKILL.md file, YAML frontmatter plus Markdown instructions, and optional scripts/, references/ and assets/ folders.
  • Same-day additions: Anthropic also added organization-wide skill management for administrators and a partner directory featuring skills from companies including Box, Canva and Notion.
  • Licensing: The specification repository publishes code under Apache 2.0 and documentation under CC-BY-4.0, and describes development as community-driven.
  • Adoption: The project's client showcase now lists 46 products that support the format, including GitHub Copilot, VS Code, Cursor, Gemini CLI, ChatGPT and Codex, goose and Kiro.

What Happened#

When Anthropic launched Agent Skills in October 2025, it described them as folders containing instructions, scripts and resources that Claude can load when they are relevant to a task. Anthropic highlighted four properties: skills are composable, portable, efficient because only what is needed gets loaded, and able to execute code where a task needs reliable, repeatable results.

The December update made the format independent of Claude. The specification at agentskills.io defines exactly what a skill is, and the GitHub repository behind it states that the format was originally developed by Anthropic, released as an open standard and is now open to contributions from the wider ecosystem. The same update added administrative controls so organizations could manage skills centrally, and a directory where partners such as Box, Canva and Notion publish ready-made skills.

The specification is short. The name field must be 1 to 64 characters of lowercase letters, digits and hyphens and must match the directory name. The description field, up to 1,024 characters, should say both what the skill does and when to use it. Optional fields cover a license, environment compatibility notes, arbitrary metadata and an experimental allowed-tools list of pre-approved tools.

Background#

The central idea is progressive disclosure. According to the specification, an agent loads only each skill's name and description at startup, roughly 100 tokens per skill. It reads the full SKILL.md body when it activates a skill, with a recommended limit of 5,000 tokens, and it opens files in scripts/, references/ or assets/ only when required. This lets an agent carry hundreds of specialized procedures without filling its context window, which is the practical limit that long system prompts and sprawling instruction files run into.

Skills sit alongside other agent standards rather than replacing them. The Model Context Protocol connects an agent to external tools and data, and AGENTS.md describes how to work in a particular repository. Skills capture reusable procedures, such as how to produce a report, run a migration or follow a release checklist, that can travel between repositories and between agents. Anthropic released the standard nine days after MCP and AGENTS.md moved to the Agentic AI Foundation, during a period when the industry was converging on shared formats for agent configuration.

Why It Matters for Developers#

For .NET teams, skills are a low-cost way to encode the procedures that coding agents most often get wrong. Because the format is plain files, skills can live in the repository, go through code review and be shared by every compatible agent the team uses, from Claude Code to GitHub Copilot. A skill for Entity Framework Core migrations might look like this:

Text
ef-core-migrations/
  SKILL.md
  references/MIGRATION-RULES.md

--- SKILL.md ---
---
name: ef-core-migrations
description: Creates and reviews EF Core migrations in this repository. Use when a task
  changes the data model, adds a DbSet, or mentions migrations or schema changes.
compatibility: Requires the .NET 10 SDK and the dotnet-ef tool
---

# EF Core migrations
1. Change the entity classes and DbContext configuration first.
2. Run: dotnet ef migrations add <Name> --project src/Data --startup-project src/Api
3. Review the generated migration against references/MIGRATION-RULES.md.
4. Never edit a migration that is already merged; add a new one instead.
5. Run dotnet test before finishing.

A few practical guidelines follow from the specification:

  • Write descriptions for the router, not the reader. Agents decide whether to activate a skill from its description, so include the keywords a task is likely to contain. Our prompt engineering guide covers the same principle for tool descriptions.
  • Keep SKILL.md short and push detail into references. The specification recommends staying under 500 lines and keeping file references one level deep.
  • Treat skills with scripts as code. Anthropic's own guidance is to use skills only from trusted sources because they can execute code. Review third-party skills as you would any dependency, and see our responsible AI guide for the prompt-injection risks of instructions an agent follows.
  • Combine skills with tools. A skill can tell an agent when and how to call an MCP server you built with the C# SDK, so procedure and capability stay separate.

For a wider view of how skills fit into agent-assisted work on .NET codebases, see our guide to AI-assisted .NET development.

What's Next#

Adoption has been the main story since December. The project's client showcase lists 46 agent products that support the format, spanning coding agents such as Cursor, Gemini CLI, OpenAI's Codex and Kiro, editors such as VS Code with GitHub Copilot, and frameworks such as Spring AI. Anthropic has also kept extending its own use of skills. Its platform release notes record an open-source Claude API skill in April 2026, and Claude Platform on AWS, launched in May 2026, supports Agent Skills.

Open questions remain. The allowed-tools field is still marked experimental, and support varies by agent, so a skill that depends on it may behave differently across products. Distribution is also less mature than for code packages, so provenance and review remain largely the responsibility of each team.

Sources#