Daily · Article 02 · 3 min read
- Open source
- Anthropic
Anthropic donates the Agent Skills standard to a neutral foundation
Official announcement dated October 22 · rechecked after 06:00 JST October 23
Summary
Anthropic announced on October 22 that it is donating Agent Skills to the Linux Foundation’s Agentic AI Foundation. The format packages procedures and resources for recurring work. Moving between tools still requires checking integrations, permissions and execution environments.
If each monthly report starts with the same instructions about matching periods and using the company template, you can keep that procedure in files. Agent Skills provides a common format for handing this knowledge to AI.
Key facts
Anthropic, the company behind Claude, announced the donation of Agent Skills to the Linux Foundation’s Agentic AI Foundation (AAIF) on October 22. The name, logo, GitHub organization and agentskills.io domain are included; using or contributing to the standard doesn't require joining. Versioned releases, a formal proposal process and security guidance are proposed roadmap items. Donation announcement

1. A handover folder for an AI worker
An agent is an AI program that reads information and uses tools to carry out work. A skill is a folder explaining how to do a particular task. Rather than retraining a model, it supplies procedures and resources to read when work begins.
The minimum is a SKILL.md file with a name, description and instructions. It can also contain code, references and templates. Format specification
Imagine a fictional shop creating a monthly sales-report skill.
| Material | Example contents chosen by the shop |
|---|---|
| SKILL.md | Use for monthly reports; compare matching periods and flag missing data |
| Reference | Definitions of sales, cancellations and refunds |
| Template | Spaces for titles, units and sources |
| Optional script | Code that totals the supplied spreadsheet |
The cat asks Q&A
- The researcher
Is that like saving a prompt I use often?
Pocket NotebookReusing instructions is similar. A skill can keep supporting documents, templates and code together too. What gets read or executed depends on the agent's capabilities and permissions.
2. Reading the relevant instructions when needed
Loading every long skill document every time would add unrelated material. The standard describes discovering relevance by name and description, reading the instructions when needed, then opening supporting resources. How skills work
When a new template arrives, the shop can update the file and instructions together. Writing “compare matching periods” doesn't ensure every result is correct. Example data can reveal whether periods, units and missing-data labels are actually handled as intended.
The cat asks Q&A
- The researcher
Does a good name guarantee the right skill is selected?
Pocket NotebookNames and descriptions are relevance clues. Vague or overlapping descriptions can lead to unexpected choices. Try representative requests to see whether the skill is used appropriately.
3. What to check after moving the folder
A common format makes instructions easier to share. But writing “read the sales system” doesn't give a new tool its integration or account permissions. MCP, the Model Context Protocol, connects AI applications to external data and functions. An MCP connection can provide tools needed by a skill's procedure. MCP introduction
| What can be shared | What to check in the new environment |
|---|---|
| Steps and definitions | How the agent loads skills |
| Report template | Reading and saving capabilities |
| Calculation code | Required programs and versions |
| Sales-system lookup procedure | Actual integration, login and read permissions |
The cat asks Q&A
- The researcher
Will another AI produce the same result as Claude from that skill?
Pocket NotebookSharing instructions doesn't guarantee identical results. Models, tools and paths can differ. Compare outputs on the same example data and fix mismatches.
4. What a neutral home changes
This announcement moves the standard toward shared governance outside one company. It doesn't certify every skill or change all skills to one license. The specification lets individual skills declare licenses and environment requirements.
Before adopting an outside skill, read its instructions, license, bundled code and required permissions. Skills can contain executable code, so treating every folder as merely a document misses part of what it can do. How code is bundled
The cat asks Q&A
- The researcher
Must we publish a skill containing our company rules?
Pocket NotebookUsing the format doesn't publish the folder. You can keep it inside the company. Before sharing externally, check for customer data and internal connection details.
To close
Knowledge that survives until next month
Keeping report definitions and checking procedures can reduce repeated explanations when the person doing the work changes. Tests of whether new tools follow those instructions make the knowledge more useful to reuse.
Sources
- Donation announcementanthropic.com
- Format specificationagentskills.io
- How skills workagentskills.io
- MCP introductionmodelcontextprotocol.io
- How code is bundledagentskills.io