Reuse a playbook
Save a way of working as a Skill, then let each agent of your organization find it, run it, change it and delete it from your AI client.
A Skill holds a playbook: the instructions that an agent follows for one ad-ops task. The playbook tells the agent which AdCrunch tools to call, in which order, and what to report. You write a Skill once. Then each agent that connects to your organization can run it in the same way.
Your Skills belong to your organization. The console lists them on the Skills page, and the tools of this job have names that start with skill_. A new organization has no Skills.
Save a Skill
Do a task with your agent once. Then ask the agent to save it as a Skill.
Save what we just did as a Skill called “Weekly budget review”. I want to run it every Monday.
The agent writes the playbook and a description, and it calls skill_create. The description tells an agent when to use the Skill, so a good description says when and why.
AdCrunch makes a slug from the name: weekly-budget-review. The slug is the handle of the Skill, and each agent uses it to find the Skill. If another Skill of your organization already has that slug, AdCrunch saves nothing, and the agent asks you for a different name. AdCrunch never adds a number to a slug to make it free.
Find a Skill
Which Skills do we have?
The agent calls skill_list. You see the name and the description of each Skill, newest first. The list does not hold the playbooks, so it stays short when a playbook is long. When your organization has more Skills than fit on one page, the agent asks for the next page.
Run a Skill
Run the weekly budget review.
The agent calls skill_get with the slug, and it reads the playbook. Then it does what the playbook says: it calls the AdCrunch tools that the playbook names, and it reports the result to you.
A Skill gives the agent no new access. Each tool that the playbook names checks its own scope, and the guardrails of each tool apply. For example, a campaign that an agent creates always arrives paused. Auth & scopes lists each scope and each guardrail.
You can also start a Skill yourself, from the prompt menu of your client. A client that shows MCP prompts lists each Skill there, often behind a / slash menu. When you choose a Skill, the client sends its playbook as your message. Prompts tells you more.
Change a Skill
Add a rule to the weekly budget review: end with a summary of one line.
The agent reads the Skill with skill_get, to get its revision. Then it calls skill_update with that revision and the parts that change. The parts that you do not name stay the same. The revision increases by one.
The revision protects a change that somebody else made. A teammate can edit the same Skill in the console, or another agent can edit it. If the Skill changed after your agent read it, the update fails with revision_mismatch, and AdCrunch changes nothing. The failure tells the agent to read the Skill again and to apply your change to the new version.
Delete a Skill
Delete the Black Friday audit Skill.
The agent reads the Skill to get its revision, and it calls skill_delete. AdCrunch marks this tool as destructive, so your client can ask you to confirm first. The same revision check applies: if the Skill changed after the agent read it, AdCrunch deletes nothing.
After the delete, the Skill is not in the list, in the prompt menu, or in the console. Its slug is free for a new Skill. You cannot restore a deleted Skill from the console or from your client.
What the agent needs
This job uses two scopes:
skill:readto list and read Skills, and to show them in the prompt menu.skill:writeto save, change and delete Skills.
A tool call without its scope fails with the code forbidden. The failure comes back to the agent as a tool result, and not as an HTTP error. Its message names the missing scope, so the agent can tell you what to do: authorize AdCrunch again from the settings of your client. Errors shows the shape of each failure.