Skills

What it's about

A skill is a set of instructions for the AI: a SKILL.md file that describes how to do a specific task – optionally with scripts, templates and reference files. The AI loads a skill on its own as soon as a task matches its description. You can also pick it for a single message or attach it to an assistant.

Skills live in their own library: in the sidebar you'll find the Skills entry, right below the Prompt library. A skill doesn't belong to any particular assistant – you create it once and use it in any chat and in as many assistants as you like. Skills are available in every workspace.

The format is the open Agent Skills↗ standard. Skills you wrote for other AI tools can therefore usually be uploaded as they are.

When is a skill worth it?

A skill pays off when the AI should do a task the same way every time and you don't want to explain it again and again:

• Fixed procedures: e.g. "How we check an incoming invoice", with your team's steps and rules
• Format and style: quotes, minutes or texts in your structure and tone
• Templates and references: a Word template, a price list or a glossary the AI looks up when needed
• Exact logic via a script: calculations, conversions or file generation that a script handles more reliably than the model

How the AI decides when a skill applies

Of all the skills available in a chat, the AI first sees only their name and description. Only when a task matches a description does it read the full SKILL.md and follow it. The chat then shows the step Read skill instructions, and below the answer you see which skill was used – e.g. Used skill invoice-check (automatic).

The description is therefore the trigger. It should say what the skill does and when it should apply – in the words someone would use in their request. It also helps to say when it should not apply.

# Too vague – the AI doesn't know when to load the skill
description: Helps with reports.
 
# Good – purpose, trigger and boundary
description: >-
  Creates the monthly sales report in the company layout from a CSV
  or Excel file. Use when someone asks for the monthly report, sales
  figures or a revenue overview. Not for single figure questions or
  presentations.

Using skills in the chat

There are three ways a skill comes into play in a chat:

Which skills are available to you: your own, those shared with you, those of the assistant in the current chat, and the built-in skills, as long as their plugin is active in the chat.

Creating a skill

In the library, click Create skill in the top right. You have three options:

After creation, MeinGPT checks the skill (status Being checked). After that it's ready to use, or you see one of these notes:

• Description missing: name or description is missing from the header of the SKILL.md. Add them – without a description the AI doesn't know when to load the skill.
• Invalid: the package can't be used, e.g. because the SKILL.md is missing, a limit is exceeded or a file contains a private key. The detail view tells you why.

Whether you may create skills is set by workspace admins under Permissions (Create skills, default: everyone).

Structure of a skill

A skill is a folder with a SKILL.md and optional additional files:

monthly-report/
├── SKILL.md          ← required: description and instructions
├── scripts/          ← optional: executable code (Python or Node)
│   └── report.py
├── references/       ← optional: reference files the AI reads when needed
│   └── metrics.md
└── assets/           ← optional: templates, images and other files
    └── template.docx

The SKILL.md starts with a header (YAML frontmatter), followed by the instructions in free-form Markdown:

---
name: monthly-report
description: >-
  Creates the monthly sales report in the company layout from a CSV
  or Excel file. Use when someone asks for the monthly report, sales
  figures or a revenue overview.
---
 
# Monthly report
 
## Procedure
1. Read the file and check that the columns region, month and revenue exist.
2. ...

Keep the SKILL.md short and move long details into references/ – the AI only reads them when it needs them.

Limits per skill

Scripts and the Code Sandbox

Skills that consist of text only (instructions, references) work without the Code Sandbox.

If a skill contains a scripts/ folder or binary files (e.g. a Word template or images), it needs the Code Sandbox. You can tell by the Sandbox label on the card and Needs sandbox in the detail view. If the Code Sandbox is turned off in your workspace, it says Instructions only without sandbox: the AI can then read the instructions but not run the scripts.

When the Code Sandbox is active in the chat, MeinGPT provides a needed skill in the sandbox automatically before the AI runs code there:

• Dependencies from a requirements.txt (Python) or package.json (Node) in the skill's top folder are installed automatically.
• MeinGPT detects from the files whether a skill uses Python or Node.
• Each version of a skill sits in its own folder in the sandbox; the AI writes results to the sandbox workspace and returns them to you as a file in the chat.

Attaching skills to assistants

In the assistant editor, right below Knowledge, there is the Skills section:

• Add skill opens a search over all skills you're allowed to see.
• Create skill offers the same three options as the library; the new skill is attached to the assistant right away.
• With the Always on switch, the AI gets the skill's full instructions with every answer instead of loading them only when a task matches. Use this for rules that should always apply, such as tone or formatting requirements.
• Use the × to remove a skill from the assistant. The skill itself stays in the library.

To attach a skill to an assistant, you need edit rights on the assistant and at least Can view on the skill. Anyone who uses the assistant can use its skills in the chat without the skill having to be shared with them separately.

Detail view, editing and versions

Clicking a skill opens its detail view. At the top you see name, origin, version and – under When the skill applies – the description. Below are four tabs:

• Instructions: the SKILL.md, rendered for reading
• Files: all files of the skill. With edit rights, use Edit to change, add or delete files and apply the changes with Save as new version.
• Usage: which assistants the skill is attached to and when it was last used
• Versions: all previous versions. Select one to see the line-by-line changes compared to the version before it.

Every change creates a new version; earlier versions stay unchanged and can be compared. Via the More actions menu you can duplicate a skill (e.g. to adapt a shared skill as your own copy) or archive it. An archived skill disappears from the library, from every assistant and from the chat.

Sharing

You share skills like assistants: click Share in the detail view, search for people or teams and choose the access level:

• Can use – use it in the chat, but not view its contents
• Can view – use it and view the instructions and files
• Can edit – additionally save new versions and manage sharing

Workspace admins can also share a skill with Everyone in this workspace (only Can use or Can view there). In the library you filter using the tabs All, Yours, Shared and Plugins.

Built-in skills

MeinGPT ships skills that belong to a plugin. You find them in the library under the Plugins tab, each with the name and icon of its plugin; they are read only. A built-in skill is available to the AI as long as its plugin is active in the chat and you may use it – the same rules as for the plugin's functions apply. If a skill needs a specific function (e.g. creating drafts) and that function is switched off in the workspace, MeinGPT does not offer the skill. Built-in skills apply automatically; you cannot pick them in the input field.

The detail view says which plugin a skill belongs to. Workspace admins get from there straight to the plugin's page, and the plugin page under Plugins lists the plugin's skills.

Files and sandbox

Plugins for email, calendar, CRM and more

The skills instruct the AI to show you what is about to happen and wait for your confirmation before anything is sent, booked, changed or deleted. A skill is guidance and gives the plugin no additional rights: the rights and approvals of the plugin's functions apply unchanged, and a function switched off in the workspace stays off.

If you want to adapt a built-in skill to your needs, create your own skill – e.g. with your formatting requirements for Word documents or your tone of voice for email replies.

Migration of existing skills

Skills that used to be attached to an assistant as a .zip are moved into the library automatically:

• Each package becomes its own skill. Its owner is the person who owns the assistant.
• The skill stays attached to its assistant; the assistant keeps behaving as before.
• Everyone who was allowed to edit the assistant may also edit the skill.
• If name or description is missing from the SKILL.md header, the skill shows Description missing. It keeps working on the assistant; still add the description so the AI finds it reliably.

If an old package contained credentials (e.g. in a secrets.json), remove them and connect the system through a connector or the Bridge instead.

Limitations

• No reset: earlier versions can be viewed and compared, but not restored directly. If needed, carry the old state over into a new version via Edit.
• Scripts only with the sandbox: without an active Code Sandbox, the AI only uses a skill's instructions.
• Formatting fidelity is not guaranteed: when a skill fills a Word template, it takes over the template's styles. MeinGPT doesn't automatically check whether the result matches your brand guidelines exactly – review brand-critical documents before sending them.

Where to go next

• Assistants – how to combine skills with a role, knowledge and tools
• Code Sandbox – where a skill's scripts run
• Prompt library – if you only want to reuse a fixed text