An Agentforce at Universal Containers is working on a prompt template to generate personalized emails for product demonstration requests from…

In prompt engineering for Large Language Models (LLMs), it is crucial to clearly separate instructions from the data the model will process. Surrounding the guidelines with triple quotes (""") creates a distinct, multi-line text block. This acts as a clear delimiter, signaling to the AI that the enclosed text is a set of specific instructions or constraints it must follow. This method prevents the model from confusing the guidelines with the dynamic data (from merge fields) or the main part of the prompt, ensuring the generated output strictly adheres to the specified rules.

B. Make sure merged fields are defined.

Defining merge fields is necessary for injecting dynamic data (like opportunity details) into the prompt, but it is not the method for writing or formatting the instructions for the AI.

C. Use curly brackets {} to encapsulate instructions.

Curly brackets ({! ... }) are the specific syntax used for merge fields to reference and insert Salesforce record data. Using them for static instructions would be a syntax error.

1. Salesforce Help

Einstein Generative AI

"Prompt Template Fundamentals": This documentation outlines the core components of a prompt template

which include the instruction (what you want the LLM to do) and the context (the information it should use). It emphasizes the importance of clear and specific instructions for guiding the LLM's output. The use of delimiters like triple quotes is a standard technique to achieve this clarity. (Refer to the section on structuring prompts).

2. Salesforce Developers Blog

"Get Started with Prompt Builder for Sales Cloud"

February 20

2024: This article provides examples of prompt templates where delimiters (such as ### INSTRUCTIONS) are used to clearly structure the prompt and separate the instructional part from the data and desired output format. This principle of using delimiters is directly implemented by using triple quotes to encapsulate a block of instructions.

3. Wei

J.

et al. (2022). "Finetuned Language Models Are Zero-Shot Learners". arXiv:2109.01652 [cs.LG]: This foundational paper on instruction tuning demonstrates that LLMs perform significantly better when tasks are framed as explicit instructions. The practice of clearly delineating these instructions within a prompt

for which triple quotes are a common convention

is a direct application of this research. (See Section 2

"Instruction Tuning").