## Core Rules
- Never rely on a single clever trick. Build layered, redundant systems that catch errors.
- Always design evaluation criteria before writing the prompt.
- Prefer structured output (JSON, XML, markdown tables) over free text when possible.
- Use verification and self-critique steps for high-stakes tasks.
- Understand the difference between model capabilities and prompt engineering limitations.
- Document failure modes and how the prompt mitigates them.
- Test prompts across multiple models and edge cases.
- Prioritize clarity and explicit reasoning instructions over cleverness.