Skills
Write reusable instructions once, publish them for a team or the workspace, and let agents load them when the job calls for them.
A skill is a reusable set of instructions, not a connection. It teaches an agent how to do a job with the tools and access it already has.
Write a skill an agent can recognize
- Name and slug
- A recognizable name and stable identifier for the skill.
- Description
- A plain account of when the skill applies, so an agent can choose it for the right job.
- Instructions
- The process, constraints, and reference material the agent should follow after loading it.
Video placeholder
Create a team skill with a concrete description and short instructions, run a matching agent task, then show the agent loading the skill in the session.
Choose its scope when you create it
A team skill is available to that team automatically. Team members can create and manage skills for teams they belong to.
Admins can create a workspace-wide skill instead. It stays out of a team until an admin grants it from that team's settings.
Image placeholder
Capture Settings > Skills with the add-skill dialog open. Show the name, slug, description, scope, and instruction fields with safe demo content.
Describe the trigger, then write the method
Put selection language in the description: name the request, output, or condition that should cause an agent to load the skill. Keep the instruction body focused on what to do after that choice.
Don't store tokens, passwords, or account credentials in skill text. Use a connection or team environment variable for secret values.
Let tools and skills do different jobs
A skill can tell an agent how your team reviews an incident. The connection supplies the Datadog or Slack tools; the skill does not grant either account.
Read how connections grant toolsArchive skills agents should stop using
Archiving removes the skill from agent access. Update the skill when its method changes; archive it when the job itself no longer exists.