Skills and slash commands
Teach the agent a rule once and it applies it in every future chat. Skills are shared with your team, and slash commands are the shortcuts that create and manage them.
A conversation is not a reliable substitute for reusable team instructions. A skill is how you make something stick: a short instruction the agent reads back in future chats, shared with everyone on your team.
A skill is a name, a description (the "when to use this" trigger) and a body of markdown instructions. The agent sees only the name and the description each turn, and loads the body when a task actually matches. So a skill costs almost nothing until it is needed.
Creating a skill from chat
Type /skill followed by the rule, in any chat:
/skill always format UK phone numbers as +44 7xxx xxxxxx, never 07xxxThe agent writes the skill from your rule plus the conversation around it, saves it to your team, and confirms with the name it chose. It applies from your next message onwards, not to the chat you are in.
/skills is accepted too, and the rule can run onto the next line: press shift+Enter
and keep typing.
If you would rather approve a draft before it is saved, just ask in plain language ("save this as a skill") and the agent drafts one for you to confirm.
Slash commands
Type / in the composer to see what is available. The composer highlights a command
once it recognises it, and only lists the ones that surface can actually run:
| Command | Where | What it does |
|---|---|---|
/skill <rule> | New chat and thread | Saves the rule as a team skill. |
/compact | Inside a chat | Summarises the conversation so far to free up the context window. |
A command that cannot run as you typed it tells you why rather than sending something
you did not mean. /skill with no rule after it asks for the rule, and /compact
before a conversation exists points out that there is nothing to compact yet.
The Skills page
Skills in the sidebar lists everything your team has, plus the built-ins.
- Applied counts. Each team skill shows how often the agent has loaded it for a task, so a skill nobody triggers reads as "Not applied yet". That usually means the description does not match how people actually phrase the work: it is the trigger, so make it say when the skill applies, not what it contains.
- Enable or disable. A disabled skill stays on the page but is invisible to the agent.
- Edit. Change the name, the description or the body at any time. Changes apply to the next message anyone sends.
- Export and import. A skill downloads as a
SKILL.mdfile (YAML frontmatter plus the markdown body) and imports back the same way, so you can review one in a pull request or move it between teams.
Skills are team-wide. Members with skill-authoring permission can create one; enabled skills are available to the team's agent. Business users do not have skill-authoring permission by default.
Built-in skills
DataMaker ships skills of its own. They are read-only, always on, and appear on the Skills page marked as such:
| Skill | Applies when |
|---|---|
| CSV synthetic data | You upload CSVs and want realistic synthetic rows generated from them. |
| DataMaker data types | A template needs the right field types instead of everything as text. |
| Nested payload templates | A nested JSON or CSV payload has to become one or more linked templates. |
| Plan test data | You ask to plan, design or scope work before anything runs. |
| Save agent scenario | You ask to make the run you just did repeatable. |
| Skill creator | You ask to capture the conversation as a reusable skill. |
| Xray test from scenario | A scenario should become an Xray Cloud test. |
| Xray execution workflow | A finished run should be published as an Xray Cloud test execution. |
Your own skills sit alongside these and are applied the same way.
Writing a skill that gets used
- Write the description as a trigger. "When generating German customer addresses" beats "German address rules". The description is what the agent matches against.
- One rule per skill. A skill that covers three unrelated things matches none of them cleanly.
- Put the detail in the body. Examples, edge cases and formats belong there. It is only read when the skill actually applies, so keep it focused on the task it supports.
- Say what to do, not what to avoid. "Use +44 7xxx xxxxxx" works better than "do not use the wrong format".
See Agent mode for how a run works, and for saving a chat as a scenario when what you want repeated is a whole workflow rather than a rule.