AI & Automation
November 16, 2026

Agent Skills: The open standard behind Claude Skills and what it means for businesses

Agent Skills explained: What the open standard covers, which tools use it, and why your skills survive a tool switch. Includes a review of 71 skills.

Agent Skills: The open standard behind Claude Skills and what it means for businesses

Less manual, more automated?

In an initial consultation, let's find out where your biggest needs lie and what optimization potential you have.

If you're building skills for Claude today, you might quickly wonder: what happens to them if we switch to ChatGPT, Copilot, or another tool tomorrow? The answer is Agent Skills. The format behind Claude Skills is an open standard, and it’s now supported by over 40 AI tools.

‍

In this article, we explain what the standard covers, which tools use it, and where things still get tricky. To do this, we tested 71 skills against the specification's rules. This includes official skills from Anthropic and OpenAI, as well as the skills we use ourselves. The basis for this is the specification on agentskills.io and the documentation for the tools, as of November 4, 2026.

‍

What are Agent Skills?

Agent Skills are an open format you can use to teach AI agents new capabilities. A skill is a folder containing a file named SKILL.md, which includes the name, description, and instructions, along with optional scripts and templates. The format was developed by Anthropic, and today it is supported by over 40 tools.

‍

According to agentskills.io Anthropic originally developed the format for Claude and then released it as an open standard. It is being developed openly on GitHub, and contributions from other companies are explicitly encouraged. The goal is stated right on the homepage: build a skill once and use it in any compatible agent.

‍

Recommended Read

OpenAI relies on modular workflows too: To see how to set up reusable skills in OpenAI’s ecosystem and what to watch out for when migrating legacy GPTs, check out our guide on ChatGPT Skills.

‍

What the standard covers

The specification is intentionally lean. The only requirement is a folder with a SKILL.md file. Alongside this, there can be subfolders for scripts (executable code), references (reference materials), and assets (templates and data).

‍

The SKILL.md file begins with a header in YAML format. Two fields are mandatory, and four are optional:

‍

  • name (mandatory): Maximum 64 characters, lowercase letters, numbers, and hyphens only. The name must match the folder name.
  • description (mandatory): Maximum 1,024 characters. It describes what the skill does and when the agent should use it.
  • license: The license under which you share the skill.
  • compatibility: Maximum 500 characters regarding requirements, such as "requires Python" or "intended for Claude Code."
  • metadata: Optional additional information such as author or version.
  • allowed-tools: Tools that the skill is permitted to use without asking for confirmation. According to the specification, this field is still experimental.

‍

Below this, you'll find the actual instructions in Markdown, with no specific formatting requirements. A skill is loaded in three stages. To start, the agent only reads the name and description—about 100 tokens per skill. If a task is a good match, it loads the full instructions; we recommend keeping these under 5,000 tokens and 500 lines. Scripts and reference materials are only fetched when needed.

‍

Which tools support Agent Skills

TheThe list on agentskills.io mentions 46 tools. For businesses, the most interesting ones are: Claude and Claude Code, ChatGPT and Codex, GitHub Copilot and VS Code, Cursor, and Gemini CLI. These are joined by JetBrains Junie, Mistral Vibe, and the data platforms Snowflake and Databricks. The office assistants Microsoft 365 Copilot and Gemini in Google Workspace are not on the list.

‍

Where the tools look for your skills varies. According to their documentation:

‍

Tool Where It Finds Skills Who It's For
Claude (Web and Desktop) Upload as a ZIP under Customize > Skills Knowledge work without an IDE
Claude Code Project folder .claude/skills and ~/.claude/skills Development teams using Claude
ChatGPT and Codex ChatGPT via upload or editor; Codex in .agents/skills and ~/.agents/skills Teams on ChatGPT Business
GitHub Copilot and VS Code .github/skills, .agents/skills, and .claude/skills, plus folders in the user directory Development within the Microsoft ecosystem
Cursor .cursor/skills, .agents/skills, .claude/skills, and .codex/skills Development teams using Cursor
Gemini CLI .gemini/skills and .agents/skills Teams within the Google ecosystem

‍

The most important practical finding: the .agents/skills folder has established itself as the common denominator. Codex, Copilot, VS Code, Cursor, and Gemini CLI all read from it. Claude Code uses .claude/skills, which Copilot, VS Code, and Cursor also monitor. If you place skills in both folders, you'll reach all the developer tools in the table. For Claude and ChatGPT, you can simply upload the same folder.

‍

Our test: 71 skills checked against the standard

Whether a skill is portable depends first on its format. We therefore wrote a test script based on the specification's rules. It checks the mandatory fields, the character rules for the name, and consistency with the folder name. It also checks length limits and fields outside the standard. We tested 71 skills:

‍

  • Anthropic, 19 official skills from the skill repository: 18 are standard-compliant. The claude-api skill, of all things, has a description of 1,068 characters, putting it just over the limit. At 617 lines, it is also longer than recommended.
  • OpenAI, 44 official skills from the skill repository: All 44 are standard-compliant. Three exceed the recommendation of 500 lines.
  • 8 skills that we run locally: 7 are standard-compliant. One has a description of 1,460 characters because it contains a large collection of trigger phrases.

‍

Not a single skill used fields outside the standard. The format itself is therefore stable; 69 out of 71 skills passed the test.

‍

Another finding is more revealing. All 19 Anthropic skills and all 8 of our own skills have descriptions longer than 200 characters. For OpenAI, that applies to 32 out of 44. This is permitted under the standard. However, if you upload a skill as a ZIP file in the Claude browser interface, according to Help Center use a maximum of 200 characters. It also allows names like "Brand Guidelines," which include spaces and capital letters—something the standard doesn't permit.

‍

To be fair, there are two limitations to keep in mind. We’ve tested the format, not how every single tool behaves. Also, a skill might be formally correct but only work well in the environment it was designed for, perhaps because it relies on specific tools.

‍

Where tools deviate from the standard

The standard sets the minimum requirements. Many providers build their own features on top of that:

‍

  1. Claude Code extends skills with custom features according to its documentation, such as controlling who can invoke a skill or executing it within a sub-agent. Other tools don't support these extensions.
  2. Claude in the browser limits the description to 200 characters during upload and allows for more flexible naming than the standard.
  3. Codex lists all skills with their descriptions at startup, using a maximum of 2% of the context window. If many skills are installed, Codex truncates the descriptions first.
  4. VS Code silently fails to load skills with invalid characters in their names, as per its documentation, without providing an error message.
  5. Gemini CLI asks for your approval before activating any skill.

‍

On top of that, there’s one thing Agent Skills intentionally don't cover: access to other systems. A skill describes how a task is performed. The connection to your CRM, email, or accounting software is handled via connectors like MCP, and those are set up differently depending on the tool.

‍

Recommended Read

To see how Skills and MCP work together in practice and which setups deliver real ROI for teams, check out our guide on Claude Skills for businesses.

‍

What does this mean for your company?

First, the good news: the skills you build belong to you and will run on many different platforms. If you switch from Claude to ChatGPT or add Copilot for development, you can take your instructions with you. For companies, this is a compelling reason to package knowledge into skills rather than leaving it buried in chat histories or the settings of a single tool.

‍

To make this work, we recommend following these five rules:

‍

  1. Clean naming: Use only lowercase letters, numbers, and hyphens, matching the folder name.
  2. Concise core description: Put the most important triggers in the first 200 characters. This ensures the skill works even when uploaded to Claude or if Codex truncates the text.
  3. Keep it lean: Keep SKILL.md under 500 lines and move details into separate files under references.
  4. Label tool-specific content: If a skill requires specific features, note that in the compatibility field.
  5. Use a central folder: Manage your skills in a shared repository and deploy them from there to .claude/skills and .agents/skills.

‍

Treat skills like code: use version numbers, review them before release, and assign an owner. Since skills can contain scripts, always read through a third-party skill before using it.

‍

Recommended Read

To see how coding agents leverage skills in practice, check out our article Coding Agents Explained. For the fundamentals behind autonomous systems, read Agentic AI Simply Explained.

‍

Bottom line: Agent Skills make your AI knowledge provider-independent

Agent Skills are the open standard behind Claude Skills, and over 40 tools support them, from ChatGPT and Codex to Copilot and Cursor. In our testing, 69 out of 71 skills met the specification, so the format is ready for prime time. The differences lie in the platforms: shorter upload limits in Claude, truncated descriptions in Codex, and custom extensions in Claude Code. If you name your files, descriptions, and folders properly, you can take your skills with you when you switch tools.

‍

The best way to start: Take your most important skill and check it against the standard using the official skills-ref validation tool. Then, shorten the description so that the most important information is in the first 200 characters.

‍

If you want to know how to build skills for your team and maintain them across different tools, get in touch with us for AI consulting.

‍

Ready to scale with AI?

Get a free strategy call and discover how AI can help you automate workflows, create better content, and scale your marketing.

Book a call

‍

FAQ: Frequently asked questions about Agent Skills

What are Agent Skills?

Agent Skills are an open format for instructions that help AI agents complete specific tasks. A skill consists of a folder containing a SKILL.md file and, optionally, scripts, reference materials, and templates.

‍

Who developed the Agent Skills standard?

Anthropic developed the format for Claude and released it as an open standard. The specification is available at agentskills.io and is being developed openly on GitHub.

‍

Which tools support Agent Skills?

According to agentskills.io, 46 tools, including Claude, Claude Code, ChatGPT, Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI, JetBrains Junie, and Mistral Vibe. Microsoft 365 Copilot and Gemini in Google Workspace are not currently on the list.

‍

Do Claude Skills work in ChatGPT?

Generally yes, because both use the same SKILL.md format. Check beforehand whether the skill requires features that only Claude has, and test it with a typical task after uploading.

‍

What is the difference between Agent Skills and MCP?

A skill describes how a task is performed. MCP connects the agent to external tools and data. They complement each other: MCP provides access, and the skill explains how the agent should use it.

‍

How do I check if my skill meets the standard?

Use the official skills-ref validation tool recommended by the specification. The command skills-ref validate followed by the path to the skill folder checks the header and naming conventions.

‍

blog

Similar posts

Less manual, more automated?

Let's arrange an initial consultation to identify your greatest needs and explore potential areas for optimisation.

SLOT 01
Assigned

To achieve the best results, we work with a maximum of six companies per quarter.

SLOT 02
Assigned

To achieve the best results, we work with a maximum of six companies per quarter.

SLOT 03
Assigned

To achieve the best results, we work with a maximum of six companies per quarter.

SLOT 04
Assigned

To achieve the best results, we work with a maximum of six companies per quarter.

SLOT 05
Available

To achieve the best results, we work with a maximum of six companies per quarter.

SLOT 06
Available

To achieve the best possible results, we limit the number of companies we work with to a maximum of six per quarter.

faq

Your questions, our answers

What does bakedwith actually do?

bakedwith is a boutique agency specialising in automation and AI. We help companies reduce manual work, simplify processes and save time by creating smart, scalable workflows.

Who is bakedwith suitable for?

For teams ready to work more efficiently. Our customers come from a range of areas, including marketing, sales, HR and operations, spanning from start-ups to medium-sized enterprises.

How does a project with you work?

First, we analyse your processes and identify automation potential. Then, we develop customised workflows. This is followed by implementation, training and optimisation.

What does it cost to work with bakedwith?

As every company is different, we don't offer a single flat project rate. One-off projects start from €5,000 (excl. VAT); the exact price follows from the scope after an initial consultation. The AI and automation subscription starts from €4,000 per month (excl. VAT). First, we analyse your processes. Then, based on this analysis, we develop a clear roadmap including the required effort and budget.

What tools do you use?

We adopt a tool-agnostic approach and adapt to your existing systems and processes. It's not the tool that matters to us, but the process behind it. We integrate the solution that best fits your setup, whether it's Make, n8n, Notion, HubSpot, Pipedrive or Airtable. When it comes to intelligent workflows, text generation, or decision automation, we also use OpenAI, ChatGPT, Claude, ElevenLabs, and other specialised AI systems.

Why bakedwith and not another agency?

We come from a practical background ourselves: founders, marketers, and builders. This is precisely why we combine entrepreneurial thinking with technical skills to develop automations that help teams to progress.

Can you work with our existing tools?

Yes. We generally build upon your existing tool stack and only add new tools if they are truly necessary. Common tools include HubSpot, Pipedrive, Salesforce, Airtable, Notion, Google Sheets, Slack, Make, n8n, Zapier, OpenAI, Claude, and other AI tools.

How quickly can we get started?

After the initial consultation, we can usually quickly define the first use cases and start implementation shortly thereafter. For simple workflows, initial results can often be seen within the first few weeks. More complex systems depend on your tools, data, and internal approval processes.

Do we own the workflows you build?

Yes. Our goal is for your team to understand, use, and continue to operate the systems themselves. That's why we meticulously document the workflows and hand them over in a way that ensures the knowledge doesn't stay with us.

Do you maintain and improve workflows even after launch?

Yes. That's precisely what the subscription is for. We don't just build workflows and disappear; we continuously monitor, improve, expand, and maintain your systems.

How are you different from an in-house automation role?

Hiring takes time, and a single person rarely covers GTM strategy, automation, AI, tooling, testing, and documentation equally well. With bakedwith, you get a specialized team with proven workflow experience, without having to build everything internally from scratch.

How are you different from a freelancer?

Freelancers can be great for individual tasks. bakedwith is a better fit if you're looking for a structured partner who identifies potential, builds workflows, documents them, and continuously improves your GTM systems.

What does collaboration with bakedwith cost?

For one-off workflow projects, we offer individual pricing from €5,000 (excl. VAT). For ongoing support, we work with monthly subscription packages. The right setup depends on your goals, complexity, and the required scope of automation.

What happens during the initial consultation?

Together, we develop initial ideas, examine your current marketing and sales processes, and assess where AI and automation truly make sense. Afterwards, we prioritize the best options and decide where to begin.

Do you have any questions? Get in touch with us!