1. Write simple intent and success criteria.
Aim for simple, concise intent and success criteria fields. Digital Humans are agentic systems. They follow instructions best with shorter prompts. Writing shorter prompts will also make it much easier for you (a human) to read over your own tests! To minimize your intent length, configure the smaller details with the configuration options of Digital Humans (see next principle). To minimize your success criteria length, configure separate evaluations with custom metrics.2. Maximize determinism.
Deterministic behavior is when you can clearly define an “if this, then that”. Any deterministic behavior should try to be moved to a configuration option in the Digital Human. Examples:- For case-specific information like ID or DoB, use Digital Human traits.
- For “if agent says this, human does that”, use scripted responses.
- For human or speech-specific behavior like language, hangup behavior, and more, use behaviors.
3. Define good agent context.
One of Bluejay’s many strengths is to create test cases you can’t see. To maximize this strength, provide as much agent context as possible. Four ways to provide context:- Ensure the agent description describes what your agent (or system) does in the Agent Settings tab.
- If your agent has a workflow, ensure your agent’s workflow is properly synced or is described accurately in the agent description.
- Upload any documents related to your system or agent by creating a knowledge base for the agent in the Knowledge Bases tab.
- Verify your agent context with Jay (Bluejay AI) in the platform. Ask it about your agent and what it thinks about its context.
Resources
Configuration
Pair a strong prompt with the right voice, language, and call-flow configuration.
Use Cases
Apply these principles to real-world test scenarios.
Custom Metrics Prompting Guide
Apply the same principles to the LLM-as-a-Judge prompts that score your calls.