Tasks and skills
A skill is a SKILL.md file: natural-language instructions that the agent
follows, with a YAML header (frontmatter). An organization's skills live in
its Kadu Playbook, and they also run in the terminal of whoever
writes them (Local development).
A skill becomes a task when it declares the task: block in the
frontmatter. Without the block, it is a helper: other skills call it, but
it does not appear in the console.
| Helper skill | Task | |
|---|---|---|
| Frontmatter | name, description, user-invocable | the same, plus the task: block |
| Who calls it | another skill, or a person in Claude Code | the console, a schedule, the API or the run endpoint |
| Has a person answering during the execution | yes | no |
| Appears in the console | no | yes |
Task identifier​
Every task has a skill_slug in the format <plugin>:<skill>: the plugin
name (the workspace) and the skill's folder name. For example, the skill
finance/skills/task-baixar-pagamentos-via-boleto-pix/SKILL.md becomes the
task finance:task-baixar-pagamentos-via-boleto-pix.
The slug is the task's identity. Renaming the skill folder creates a
different task, without the previous one's history. To change the
displayed name, change task.label.
What the task declares​
The task: block defines what the console shows and how the execution runs:
| Field | Effect |
|---|---|
label, description | The task's name and introduction text in the console. |
activity | The activity under which the task appears. |
model | The task's tier: defines the price, the executor and the AI model. |
parameters_schema | The parameters accepted at run time, as JSON Schema. The console builds the form from it, and an invalid parameter is rejected before the execution starts. |
accepts_input_files, generates_output_files | Whether the task receives files at run time and whether it publishes files at the end. |
required_secrets | The secrets the execution receives. Only those listed reach the execution. |
systems, features | The systems the task uses and what it delivers, shown in the console. |
schedules, endpoint, executor, metadata | Schedules, HTTP trigger, model override and free-form metadata. |
The full format, field by field, is in
The task: block.
No person on the other side​
A task execution is automatic: nobody answers a question from the agent. When the skill hits an ambiguity or a failure, it records a pending item and moves on to the next item. Pending items stay in the console for a person to resolve later.