Salesforce lets the model write Prompt Builder drafts—but not control what becomes real

Salesforce’s Assistive Authoring system for Prompt Builder is a useful counterexample to the idea that a capable model should run the whole workflow. The model interprets an administrator’s request and drafts prompt-template content, but Salesforce keeps routing, record identity, permissions, output parsing and final action under deterministic control.

Put the graph—not the model—in charge

The workflow uses Agent Graph to route requests among three specialized paths: creating a template, refining an existing template, or clarifying a request that is unrelated or too vague. The model can signal intent, but deterministic guards decide which transitions and backend tools are allowed. A client-supplied mode also distinguishes create from refine.

That distinction matters because a linguistically plausible answer can still be operationally wrong. Salesforce identifies invented fields, altered record IDs, malformed metadata and incorrect routing as failure modes that can produce a convincing but unusable template.

Preserve identity outside the prompt

During refinement, the LLM works with stable developer names rather than opaque record identifiers. A server-side transpose layer maps generated elements back to existing template versions, inputs, providers and parent records, then reapplies the correct IDs. The model therefore never has to reproduce identifiers it could corrupt.

Structured output gets a similarly narrow contract. The server wraps template metadata in sentinel tags; the client extracts only that tagged payload and parses it as JSON. A parse failure returns an error instead of triggering a best-effort interpretation.

Turn grounding into an allowlist

Salesforce’s discovery service normalizes available data providers, attaches the correct reference names and filters fields according to the current user’s field-level readability. A relevance step reduces those validated providers to an applicable set. Generation can select from that allowlist, but it cannot assume an object or field exists.

The order is the important part: discover real data, normalize it, enforce permissions, select relevant context, and only then let the model generate. This is stronger than asking the model to avoid hallucinations in prose.

Keep drafting separate from authority

The agent can generate or refine a proposed template, but it has no tool to save, publish or execute it. Output is converted into typed editor blocks rather than injected as raw HTML, and malformed metadata is rejected. The administrator sees an accept-or-reject diff; accepting changes only the editor draft, while saving remains a separate action.

Salesforce says the system aims to reduce a process that can take an experienced administrator at least 25 minutes to under a minute. That is a vendor-reported goal, not an independent benchmark. The more transferable result is architectural: let the model handle ambiguity and language, while deterministic software owns identity, permissions, state transitions and side effects.