> ## Documentation Index
> Fetch the complete documentation index at: https://developers.autoplay.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Tool Prompts

> Descriptions that teach an AI support agent when and why to call each Autoplay MCP tool.

Attach these descriptions to each Autoplay MCP tool so the agent calls it with the right intent and interprets its result correctly. The exact field name varies by agent provider.

## `get_live_user_activity`

```text theme={null}
Use this tool to retrieve what the current user has just been doing in the product: their current page or screen, recent actions, and any in-progress or abandoned flows.

Call this tool:
- Before grounding a proactive message in what the user is actually doing.
- When the user's question is ambiguous and recent activity may clarify their intent.
- To detect signs of friction, such as repeated actions, idle time on a page, or an abandoned flow.

Do not use this tool to make assumptions about the user's goals beyond what the activity supports. If the evidence is inconclusive, ask a focused clarifying question instead of guessing. Never expose raw events or mention the tool to the user.
```

This tool answers **what the user is doing now**. It does not determine their overall onboarding priority or decide whether the agent may initiate contact.

## `get_onboarding_context`

```text theme={null}
Use this tool to retrieve where the current user sits in their onboarding journey: which milestones are complete, which are outstanding, and the next relevant workflow.

Call this tool:
- At the start of a proactive interaction, to identify the most relevant milestone to surface next.
- When acknowledging a completed action, to identify the correct next milestone and why it matters.
- To avoid re-suggesting a milestone the user has completed or explaining something they have already figured out.

Combine it with get_live_user_activity. Onboarding context shows where the user should be headed and what they have already done. Live activity shows what they are doing right now, including actions they are already correctly midway through. Check both before making a personalized recommendation.

Never expose internal stages, gates, experiments, workflow keys, or tool-output labels to the user. If degraded is true, do not assert specific progress; provide general help or ask a focused question instead.
```

This tool answers **what is genuinely next**. It does not authorize agent-initiated contact.

## `get_proactivity_criteria`

```text theme={null}
Use this tool to determine whether the current moment is appropriate for the agent to proactively initiate contact.

Call it immediately before any proactive, agent-initiated message. Do not call it before a reactive response to a user message or before help the user explicitly requested.

Respect its result strictly. Treat can_be_proactive_now as the source of truth. If it is false, do not initiate a message, even when get_live_user_activity or get_onboarding_context suggests a useful opportunity. Re-check only after the blocking condition may have changed.

Do not reveal the verdict, blocked reason, agent state, explore gate, or internal criteria to the user.
```

This tool answers **whether the agent may interrupt now**. It does not choose the recommendation by itself.

## Keep identity out of the conversation

All three tools require:

* `product_id`: the fixed Autoplay product identifier associated with the authenticated connection.
* `user_id`: the exact stable application ID from verified session context, matching the ID sent by the activity source.

Do not configure either value as information the agent should collect from the user. Never substitute an email, display name, conversation ID, or guessed value. If the agent provider cannot bind verified identity to the tool argument, resolve that integration gap before enabling personalized recommendations.

## Verify the behavior

Test three separate paths:

1. Ask a question about the current screen and confirm the agent uses live activity without exposing raw events.
2. Ask what to do next and confirm the agent does not recommend a completed or in-progress milestone.
3. Attempt an agent-initiated message when `can_be_proactive_now` is false and confirm no message is sent.

<Card title="Review the complete prompt" icon="arrow-left" href="/best-practices/system-prompt">
  See how these tool instructions fit into the provider-neutral conversation behavior.
</Card>