Testing a chat feature needs conversations, and writing forty of them by hand is nobody's afternoon. These are structurally plausible system and user messages for exercising storage, token counting and rendering — not for producing good model output.
What you can control
- System messages combine a role, a task and a configurable number of constraints, which is how most real ones are shaped.
- Pair mode produces a matching system and user turn for seeding a conversation record.
- The optional output schema gives you something to test structured-output parsing against.
- Worked examples exercise the few-shot section that many prompt templates include.
Testing ideas
- Prompt templates. Generate system and user pairs to test template rendering, escaping and variable substitution.
- Output schemas. Include an output schema to test JSON-mode parsing and validation of model responses.
- Token limits. Generate long prompts with worked examples to check truncation and token counting.
- Evaluation sets. Export a batch as JSON to seed an evaluation or regression set for an LLM feature.
What this is not
These prompts are for filling a database, not for use with a model. They are assembled from templates and will not produce useful responses.