Skills

A skill is a named instruction on how something is done. Skills live in a library of their own and can be attached to several agents.

Create a skill

Skills in the sidebar opens the skill library. There are three ways to create a skill:

• Start empty: creates a skill named Unnamed skill, without a description and without text.
• Create a new skill with AI (in the chat): starts a conversation in the chat. The chat creates the skill and shows the card Skill saved with the button Open.
• Import SKILL.md: takes over a SKILL.md file as a draft to review.

A skill from the chat is saved immediately, without an approval card. It is private. If the chat saves again in the same conversation, it overwrites the skill instead of creating a second one.

Edit and delete a skill

On a skill's detail page, the description is under When should the agent use it?. You write the text under Instruction. There you switch between Text and Preview.

Undo goes back at most one version, and only if the instruction changed. This step can itself be reversed.

If you leave the page with unsaved changes, the browser asks first. Delete skill at the bottom left deletes the skill.

Attach skills to an agent

1. Open the agent.
2. In the card Skills, click Add skill.
3. Tick the skills and click Done.

The chain symbol in the list detaches a skill from the agent again.

Above the list is the switch Use all skills. When it is on, the agent knows every switched-on skill of the workspace in the chat. When it is off, it knows only the attached ones.

Import and export SKILL.md

The import via Import SKILL.md reads Anthropic's SKILL.md format: a YAML header with name and description, followed by the text in Markdown.

The text is taken over unchanged. Next to it, the draft marks lines that work differently in lowcloud. Each note has a category: Program call, File access, Network access, Only in Claude Code or Foreign tool. The draft also warns about English text, about allowed-tools from another environment and about unknown entries in the file header.

The skill is saved only when you click Apply. If you drag several .md files at once, only the first is processed.

The export downloads the skill as a SKILL.md file. The file name is in lower case with hyphens. The exported file runs in Claude Code.

How an agent uses a skill

A skill consists of a name, a description and an instruction text in Markdown.

The agent's prompt contains only an index: the name and description of every skill. The agent fetches the full text itself when the description fits the task. The index is part of every request and costs tokens.

Skill names are not unique. If two skills have the same name, the model gets both texts and decides for itself.

Skill status

• Draft: a skill without a description. It is saved, but delivered neither in the chat nor in a run, even if it is attached.
• Off: the switch on the skill card is off. The skill is then not in the chat index, cannot be read, does not run in any run and is missing from the catalog of the build chat. Links stay in place and carry the badge Off. Switching it on restores everything.

If the index gets too large, switch skills off. lowcloud does not shorten the index by itself.

Visibility

The chip on the detail page sets a skill from Private to Workspace. It is then in the library of the workspace. Back on Private, only the author sees it.

A private skill cannot be attached to an agent in the catalog, nor to the base chat. Conversely, a skill cannot be made private as long as an agent or workflow that the team sees uses it. The message names them.

The section Used by on a skill's detail page shows where the skill is attached.

Limits

Permissions

Anyone who may build (the role admin, or the former role developer) can create, change, switch on or off and delete skills.

People with the role member cannot read the skill library or the skill texts. Their sidebar has no entry Skills, and opening the address directly redirects them to the list Agents.

What applies in a partner's client workspace is described under Partner.