Key topics
- Instruction hierarchy: system policies vs user requests vs retrieved context.
- Structured outputs and schema adherence patterns.
- Few-shot examples for edge cases and error handling.
- Refusal and escalation language that is consistent and safe.
Common pitfalls
- Prompts that try to encode business logic instead of using validation code.
- Mixing untrusted retrieved text with policies.
- Overly long prompts that increase cost and reduce accuracy.
- No versioning or tests for prompts.
Recommended practices
- Treat prompts as versioned artifacts with change control.
- Prefer explicit constraints and checklists over long narratives.
- Use schemas and validators rather than ‘please be correct’.
- Measure prompt changes via golden-set regression.
This page is intended to be actionable for engineering teams. For platform-specific details, cross-reference /platform/agents, /platform/orchestration, and /platform/knowledge.