Use AI Assist in Prompt Builder (Beta)
Use AI Assist to create a structured, Jinja-based prompt template from a natural language description. AI Assist is powered by an Agentforce agent that pulls in relevant fields from your org and structures prompts using role, task, and audience formatting.
Required Editions
| Available in: Lightning Experience |
| Available in: Enterprise, Performance, and Unlimited Editions with the Einstein for Platform, or Einstein or Agentforce for Sales or Service add-on, or Agentforce Foundations |
When you create a new prompt template, you can choose to draft it with AI Assist instead of building it from scratch. Describe what you want your prompt to accomplish in natural language, and AI Assist creates a structured, Jinja-based template with recommended inputs and grounding data. You can add resources such as objects directly in the composer, and AI Assist automatically creates the corresponding template inputs.
After AI Assist creates a draft, the template appears in the canvas with visual indicators that show what was added. You can then provide additional instructions to customize the draft before accepting the final version.
Understand How AI Assist Works
AI Assist is powered by an Agentforce agent built into Prompt Builder. The agent interprets your natural language description and creates a structured Jinja-based template that includes:
- Role, task, and audience formatting
- Recommended merge field inputs based on the resources you specify
- Grounding data recommendations relevant to your described use case
After the draft is created, review it in the canvas. Visual indicators show additions in green and removals in red. You can customize the draft by providing additional instructions — for example, Make the tone casual — and review the updated version before accepting.
Accept and Discard Changes
After reviewing the drafted or customized template, use the controls at the bottom of the canvas to accept or discard all changes. Accepting applies the draft to your template. Discarding returns the template to its previous state.
You can customize and accept changes multiple times before saving the final template.
