PromptTemplateActions

Overview

Learn how to invoke Salesforce Prompt Templates from an Agent Script agent as an action. This recipe demonstrates how to pass prompt inputs (like an email address) into a prompt template and return the generated response back to the user, while keeping the agent logic simple and repeatable.

Agent Flow

%%{init: {'theme':'neutral'}}%%
graph TD
    A[Start] --> B[Topic Selector]
    B --> C[Schedule Generation Topic]
    C --> D[Ask for email address]
    D --> E[Run Generate_Personalized_Schedule action]
    E --> F[Prompt Template executes + returns promptResponse]
    F --> G[Send schedule to user]

Key Concepts

• Prompt Template action targets: Use generatePromptResponse://<PromptTemplateDeveloperName> to execute a prompt template and return a promptResponse string.
• Prompt Template inputs: Prompt inputs are passed using the Input:<apiName> reference name (for example, Input:email).
• Grounded data via data providers: Prompt templates can pull grounded data via data providers (for example, an Apex data provider referenced from the template).
• Topic-scoped execution: Keep the “ask for missing input” + “run the action” logic inside a focused topic so the agent stays predictable.

How It Works

1. Create (or reference) a Prompt Template

This recipe includes a prompt template asset named Generate_Personalized_Schedule (genAiPromptTemplates/Generate_Personalized_Schedule.genAiPromptTemplate-meta.xml). The template:

• Accepts an email input (referenced as Input:email)
• Uses a data provider (Apex) to retrieve grounded, relevant activities for that guest
• Produces a single text response that the agent can display to the user

2. Define an Agent Script action that targets the template

In the agent file (aiAuthoringBundles/PromptTemplateActions/PromptTemplateActions.agent), the action is defined with:

• A required input named "Input:email" (matching the prompt template input reference name)
• An output named promptResponse (the standard response field returned by the prompt execution)
• A target pointing at the prompt template by developer name using generatePromptResponse://...

This indirection is useful because the agent can treat the prompt template like any other action: inputs in, outputs out.

3. Ask for required input, then run the action

The schedule_generation topic is responsible for collecting the email address. Once the user provides it, the topic runs the action and uses the returned promptResponse as the content of the next message.

This pattern keeps the conversation clean:

• The agent only asks for the one required input it truly needs
• The prompt template stays responsible for the “heavy lifting” (content generation and grounded context)

Key Code Snippets

Action Definition

actions:
  Generate_Personalized_Schedule:
    description: "Generate a personalized schedule with a prompt template."
    inputs:
      # Prompt Template inputs follow the pattern: "Input:<Field API Name>"
      "Input:email": string
        description: "User's email address"
        is_required: True
    outputs:
      # Prompt Template output is always a promptResponse string
      promptResponse: string
        description: "The prompt response generated by the action based on the specified prompt and input."
        is_used_by_planner: True
        is_displayable: True
    target: "generatePromptResponse://Generate_Personalized_Schedule"

Action Usage (Topic Instructions)

| When a user asks for help scheduling their day or creating a personalized schedule:
| 1. Ask for their email address
| 2. When you have the email address, run {!@actions.Generate_Personalized_Schedule}
| 3. Output the personalized schedule in a message

Try It Out

Example Interaction

Agent: Hi! I can generate personalized schedules using Prompt Templates.

User: Generate a personalized schedule of experiences for me.

Agent: To create a personalized schedule for you, please provide your email address.

User: alex.reed@xample.com

Agent: Here is your personalized schedule of experiences ...

What's Next

• AdvancedInputBindings: Explore richer ways to bind variables and inputs into actions.
• ActionDefinitions: Learn how to define and document actions the planner can select automatically.
• ActionCallbacks: Add post-processing logic (validation, formatting, side effects) after an action runs.