Skill

A skill is a saved procedure your agent invokes. Like Claude Skills or custom GPTs, family-shaped. Reusable, parameterized, auditable. A skill captures what your agent does: the prompt, the tool list, the inputs: so it can re-run with new context tomorrow without re-prompting.

Skills are tagged by where they make sense to run, not who runs them. generic skills only touch Sprout and can run anywhere; home_agent skills reach outside Sprout (school portal, your calendar, your folders) and only make sense in your home agent's context.

Where you'll see this

In the wild: Save it as a skill

A worked example

A skill that refreshes today's Daily check-in to reflect what actually happened at school. Tagged home_agent because it reads the school portal before updating the task.

Shell

skill.write({
name: "Refresh today's check-in",
description: "Pull today's school events and rewrite the check-in chat to match.",
category: "home_agent",
prompt: `For {{input.child_name}} ({{input.child_id}}) on {{input.today}}:
1. Read today's events for this kid from the school portal connector.
2. task.list({ assignedChildIds:[{{input.child_id}}], status:'open' });
find the open task named 'Daily check-in'.
3. Rewrite conversationSpec.guidance to fit today specifically.
4. task.update(taskId, { conversationSpec: { goalType:'share', guidance: <rewrite> } }).`,
handsReferenced: ["task_list", "task_update"],
inputVariables: [
{ name: "child_name" },
{ name: "child_id" },
{ name: "today" }
]
})
# Then each morning:
skill.invoke({ skillId: "<skill>", input: { child_name, child_id, today } })

The Save it as a skill page walks this top to bottom.

Shape

name · description: what the skill is and what it does, in the parent's language.
category , generic (default) or home_agent. The audience fork.
prompt: the procedure your agent reads on invoke. Templates use {{input.*}} placeholders; child names verbatim are refused with PII_IN_PROMPT.
handsReferenced: declared tools the procedure uses. Advisory, not an allowlist; biases attention.
inputVariables: what your agent passes on each invoke.
kidCallable: flag. Whether the kid app can launch this directly. Extra scrutiny if true.
canvasIds: optional. Canvases linked to the skill. skill.update uses addCanvasIds / removeCanvasIds (incremental, not full replacement).

Lifecycle

The skill lifecycle is flat: a skill exists or it's archived. There is no separate "activate" beat: invocation records its own usage event, and skill.get surfaces the most recent timestamp as lastTriggeredAt. Archived skills can't be invoked or scheduled; create a fresh one to reuse the pattern.

When to use it (and when not to)

Save as a skill when:

You'll want the same procedure tomorrow with different inputs.
The content has to change between runs (today's school day, this week's themes).
You want an audit trail and library entry, not a one-off shell command.

Stick with a standalone task when:

The content is the same every time; the schedule does the work.
It's truly one-off and you don't expect a sibling pattern.

Common errors

BAD_INPUT · PII_IN_PROMPT: a real child's name appeared verbatim in prompt or description. Use {{input.child_name}}.
BAD_INPUT · SPEC_HASH_MISMATCH: your commit input changed between dry-run and commit. Re-dry-run.
BAD_INPUT · SCHEDULER_AS_RUN_SKILL: passed a scheduler-shaped skill as the run skill of a heartbeat.
PERMISSION_DENIED: invoking or heartbeating a home_agent skill from Sprout-runtime. By design.

Related tools

skill.write: author. Supports dryRun + specHash.
skill.update: partial patch. Conditional gates: name/desc-only skips safety + policy.
skill.invoke: your agent renders the skill into the conversation. Not kid delivery.
skill.list / skill.get: read shape; canvas metadata denormalized on get.
skill.post_result: explicit publish to the family inbox.

Roadmap for this noun

Soon Cross-family skill library: subscribe to skills other families' agents publish. Metadata review becomes a hard gate.
Soon Granular consent: per-skill parent approval at first invoke, not just at OAuth time.

Save it as a skill Task How safety works

Was this page helpful?