You do not need to turn skills on. In every run, the agent has every skill you can access: your private skills, skills shared with your teams or organization, and System skills. It reads each skill’s name and description and uses a skill when your task matches it. You can also point the agent to a specific skill. Name the skill in your prompt, or add it from the / menu in the composer.
Use cases
- Ask for a review and let the agent apply your team’s code-review skill.
- Make sure the agent follows your api-conventions skill when it adds an endpoint.
- Combine skills in one message, for example api-conventions and release-notes.
- Apply a skill in a follow-up message after the agent has made changes.
How the agent uses skills
- It sees the list. At the start of each run, the agent gets the name and description of every skill available to you.
- It matches your task. When your request fits a skill’s description, the agent opens the skill.
- It follows the instructions. The agent reads SKILL.md and follows its steps. It reads reference files and runs included scripts when the instructions call for them. You see this in the thread’s live progress, for example as a file read of the skill’s SKILL.md.
Name a skill in your prompt
Naming the skill is the most reliable way to make the agent use it.
-
Open a thread, or select New task.
-
In the prompt box, name the skill and say what you want:
Use the code-review skill to review the changes on this branch compared with main. Report findings only. Do not change any files. -
Press Enter to send. Use the skill’s name as the agent sees it: lowercase words joined by hyphens. For example, a skill named Code Review is code-review to the agent. To use more than one skill, name each one:
Add a GET /orders/{id} endpoint. Follow the api-conventions skill,
then use the release-notes skill to draft a changelog entry for it.Add skills from the / menu
The / menu lets you find a skill by name or description and add it to the composer as a badge.
- In an empty prompt box, type /, or select Quick actions. A menu opens with Skills and Files.
- Select Skills.
- Type in Search Skills… to filter the list by name or description.
- Select a skill. Use ↑ and ↓ to move through the list and Enter to select. A check mark appears next to the skill, and the skill appears as a badge above the prompt box.
- Select more skills if you need them. The menu stays open, and the Skills item shows how many you selected.
- Press Esc or ← to go back to the first menu, and Esc again to close it.
- Type your message and send it. To remove a skill, select ✕ on its badge, or select the skill again in the menu.
Name the skills in your message too, for example “Use the code-review skill to …”. That makes sure the agent applies them.
If no skill matches your search, the list shows No skills matching “{your search}”. If you have no skills, it shows No skills available. To add one, see Create and manage skills.
Best practices
- Write the request, not only the skill name. Say what you want done; the skill says how to do it.
- Name the skill when it must apply. The agent picks skills on its own, but naming one removes any doubt.
- Use one skill per concern. Combine small skills in one message rather than writing one skill for everything.
- Check the result against the skill. If the agent skips a step, reply in the same thread, for example: Follow step 3 of the code-review skill and check the tests too.
- Improve the skill, not every prompt. If you correct the agent the same way twice, add that rule to the skill.
Troubleshooting
- The agent ignored a skill: name the skill in your prompt. If it still does not apply, make the skill’s description more specific about when to use it.
- The agent used a skill you did not want: say so in your prompt, for example Do not use the code-review skill for this task. If it happens often, narrow the skill’s description.
- A skill is not in the / menu: for a repository skill, check that its
SKILL.mdis on the branch the task uses. For a plugin skill, check the Skills tab on the Customize page. If it was added recently, reload the page. - The agent used an old version of a skill: skill changes apply from the next run. Send a new message to start a new run.