A prompt is an instruction or reusable template that shapes a model interaction. System prompts can define an assistant’s tone, boundaries, and response behavior. section 3.1.3 section 6.4.3
Keep production prompts in a prompt registry rather than copying them into workflow code. Naming and versioning prompts supports reuse, testing, and rollback without redeploying the application. section 3.1.3 section 10.2
A managed prompt workflow can:
- Store a prompt under a name with metadata such as author, reviewer, and tags. section 3.1.3
- Retrieve it by name when constructing a model request. section 3.1.3
- Compare prompt variations against the same evaluation dataset and expected answers. section 2.1.6
Listing 3.3 Prompt registration request and response
Request:
name: "patient-intake-v2" #A
content: "You are a helpful patient
➥intake assistant for HealthFirst Medical..."
metadata: { #B
author: "sarah@healthfirst.com",
reviewed_by: "legal@healthfirst.com",
tags: ["production", "patient-facing"]
}
Response:
name: "patient-intake-v2"
version: 1 #C
created_at: "2024-01-15T10:30:00Z"
Shows a prompt registration request with content and metadata, alongside a versioned response. section 3.1.3
Testing prompt variations
A patient intake assistant can compare formal and conversational system prompts against the same realistic questions before choosing one for deployment. section 2.1.6































































