Anthropic transfers Agent Skills standard to Linux Foundation

Anthropic has transferred the Agent Skills standard to the Linux Foundation, the same foundation that previously received MCP.

Since the standard launched in October last year, 1 million skills have been published and installed almost 280 million times. Today, Agent Skills are supported by more than 70 AI agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code.

Over the past year, a unified approach to creating skills has emerged. The official guide to skill development has also been donated to the foundation.

Six of the most useful rules from the guide:

1. First complete the task together with the agent. Only then turn the steps that worked, your corrections, input and output formats, and additional context into a skill. If a model writes a skill based only on general knowledge, the result will be filled with useless advice such as “handle errors properly”.

2. Review every line. Ask yourself whether the agent would make a mistake without this instruction. If not, remove it.

3. Pay special attention to the skill description. At startup, the agent reads only the name and description, so the description determines whether the skill is used at all. Phrase it as “Use this skill when...”, make it assertive enough, list situations users may not mention directly, and keep it within 1,024 characters.

4. Do not overload SKILL.md. Limit it to 500 lines and 5,000 tokens. Move details to the references/ directory, specifying when the agent should read each file.

5. Separate strict instructions from recommendations. For actions where mistakes are especially easy to make, provide exact commands. Where independent judgment is required, explain the reasons for the choice. If a task can be completed in multiple ways, offer one default option.

6. Maintain a list of common mistakes (gotchas). In many skills, this turns out to be the most useful part. Every time you have to correct an agent mistake, add a corresponding note to the list.

Immediately after the agent completes a task you often assign to it, you can send the following example prompt to Claude Code in the same session to save the successful approach as a standard skill.

Example prompt:

Following the Agent Skills Best Practices guide, turn the task we just completed into a skill.

Use only the steps that worked in this session, my corrections, and related decisions from CLAUDE.md, AGENTS.md, and Git history. Do not fill gaps with general knowledge.

Write the description in the format “Use this skill when...”. Include situations I did not mention directly but where the skill should activate. The description must not exceed 1,024 characters.

Limit SKILL.md to 500 lines. Move detailed information to the references/ directory, specifying when the agent should consult each file.

Create a separate section for common mistakes and gotchas.

Use only lowercase Latin letters, digits, and hyphens for the skill name. The name must match the directory name.

Finally, create 2–3 real test scenarios in evals/evals.json, specifying the prompt and expected result for each. Run all tests, show their results, and describe the actions taken.