A Platform Administrator is creating a new action instruction for an agent. This action, named createCase, is designed to generate a new Salesforce…

According to best practices, action instructions for an AI agent or copilot should be clear, descriptive, and focused on the action's capability and the user's intent. Option B excels by clearly stating what the action does ("create a new Salesforce Case record"), its specific goal ("document a customer’s issue"), and the precise user intent that should trigger it ("create a formal record of their problem or question"). This three-part structure provides the underlying Large Language Model (LLM) with the necessary context to accurately select and use the action at the appropriate time, distinguishing it from other potential actions.

A. This option is less effective because its trigger, "when the user wants to create a case," is circular and less descriptive of the underlying user intent than the phrasing in option B.

C. This instruction is too broad. Stating the action is for "any type of customer inquiry" could lead the agent to create cases unnecessarily, as not all inquiries require a formal case record.

D. This option is flawed because it includes irrelevant technical implementation details ("code snippet," "runs in the background") and incorrectly states the purpose is to "solve the customer’s issue," when it only logs the issue.

1. Salesforce Help Documentation, "Create Einstein Copilot Actions": This official guide outlines the best practices for writing Action Instructions. It emphasizes being specific and descriptive about what the action does and when to use it. The recommended structure is to state the capability and the triggering user intent. For example, for a summary action: "Use this action to summarize an Account record. Use this action when a user asks for a summary, a brief, or an overview of an account." Option B most closely follows this prescribed, intent-focused format.

Source: Salesforce Help Portal, Document ID: sf.copilotactionscreate.

2. Salesforce Developer Documentation, "Einstein Bots Developer Cookbook": While focused on Bots, the principles for describing actions to an NLU/LLM engine are the same. The documentation consistently advises describing the business purpose and outcome of an action in plain language, rather than its technical implementation. Option D violates this by mentioning code snippets, while Option B correctly describes the business purpose.

Source: Salesforce Developers, Einstein Bots Documentation, Section: "Dialog Actions".

3. Stanford University, CS224N: NLP with Deep Learning, Lecture Notes on In-Context Learning: Course materials on prompting large language models emphasize the importance of providing clear, unambiguous instructions with context and examples of intent. Option B provides the richest context ("document a customer's issue") and the clearest intent ("create a formal record"), which aligns with academic best practices for guiding LLMs.

Source: Stanford University, CS224N Course Materials, Winter 2023, Lecture on Prompting and In-Context Learning.