Skip to main content
A skill is a reusable set of instructions and supporting files. Install it for the workspace when every agent should have it, including agents created later. Use Agent skills for an individual agent.

Open Workspace Skills

Open Workspace Settings → Workspace Skills. Depending on the navigation layout, the workspace Skills entry opens the same browser. You need permission to manage skills to install, remove, create, import, or edit them. With view-only access, you can inspect skills and ask a workspace admin to make changes. A skill does not provide an external account, credential, tool permission, or model connection by itself.

Find a skill

The browser has two main tabs: Search by name or description. Use For to filter by role or built-in skills, and From to filter by AgentDesk managed or User managed. Clear the search or choose Everyone and Any source to broaden the results. Expand long descriptions with Show more. AgentDesk managed skills stay in sync with AgentDesk. User managed skills can have editable installed copies. Needs attention means the skill is not currently usable; review its files or reported issue before trying to install it.

Install for everyone

  1. Open All Skills and find the skill.
  2. Read its description and choose View to inspect its files if needed.
  3. Choose Install.
  4. Wait for confirmation, then check Installed Skills for the Installed badge.
Every agent inherits a workspace installation. You do not need to install it separately on each agent.

Edit an installed skill

For an installed User managed skill, choose Edit or open its name.
  1. Select a file tab, usually starting with SKILL.md. Use the editor’s Markdown preview to review the rendered instructions.
  2. Read the scope hint: this is the workspace copy.
  3. Change the text and choose Save for the selected file.
  4. Wait for the saved confirmation. Save each changed file separately before closing.
The editor shows text files that can be displayed. Each file’s availability and editability determine whether it can be changed. Keep the skill’s name in the SKILL.md frontmatter unchanged; changing the instructions or description does not rename the skill. Workspace edits affect agents that use the workspace copy. An agent’s own override remains separate. Use Agent skills when a change should affect only one agent. Reset, when available, replaces the selected file with its catalog version. Review the confirmation before proceeding. It restores the catalog baseline, rather than a previous saved edit. The catalog itself is read-only in this file editor. If another change happened since you opened the file, the editor reports a conflict. Keep a copy of your intended edits, use Reload file, compare the current content, and apply the change again. Closing with unsaved changes offers Keep editing or Discard.

Customize an AgentDesk managed skill

AgentDesk managed files are read-only. Open the skill and choose Clone to edit:
  • Clone only creates an editable User managed skill in the catalog. Install the new skill where you want to use it.
  • Clone and use here creates the editable copy and uses it in this workspace scope in place of the original. The original remains in the catalog.
Confirm the choice, then open the editable installed copy. Check the workspace and affected agent skill lists to understand which copy they use.

Create a skill

  1. In All Skills, choose New skill.
  2. Fill in the SKILL.md template. Give it a unique name and a description that explains what it does and when to use it.
  3. Write clear instructions, inputs, steps, and expected output.
  4. Choose Create skill and wait for confirmation.
  5. Find it in All Skills, then install it for the workspace or a particular agent.
For example:
Creating a catalog entry and choosing where agents use it are separate steps.

Import skills

In All Skills, choose Import skill. The same importer is available from an agent’s skill browser.
  1. Choose GitHub repository and enter a Public GitHub repository URL, or choose Archive upload and select a .skill or .zip file.
  2. Choose Preview. Review detected candidates, validation errors, and warnings.
  3. Select the valid candidates you want. Invalid or unavailable candidates cannot be selected.
  4. Review Duplicate behavior before choosing Install selected.
  5. Wait for the server-confirmed result. Return to All Skills, then choose Install in the workspace or agent browser where you want to use it.
The importer uses a public GitHub URL; it does not offer a private-repository token field. If SKILL.md is in a subfolder, use a GitHub URL pointing to that skill folder. The GitHub preview reads the detected SKILL.md; use an archive when you need to include supporting files. An upload must contain a valid skill structure, including SKILL.md. Choose replacement only after reviewing the existing version and its users. A preview confirms what was detected; it does not confirm installation. Install selected imports candidates into the catalog. It does not choose a workspace or agent installation. Install the catalog entry in the intended browser, then inspect Installed Skills to confirm its scope.

Remove or archive

Choose Remove on an installed workspace skill and confirm. This removes the workspace installation. Agents that rely on it lose that inherited copy; separately installed agent copies remain distinct. Archive is a catalog action offered for eligible User managed skills when you open the catalog version. Confirming hides the skill from the catalog and prevents new installations. It is different from removing an installation from one scope.

Refresh and resolve problems

Use Reload Skills after files or installations change outside this browser. With management permission, this action also reconciles the catalog with installed skill files. If a skill cannot be installed, inspect Needs attention, clear filters, and review validation errors. If an agent reports Missing files, follow Agent skills. If a save fails, read the error and check your permission, file size, and whether the file changed while the editor was open.