The prompting playbook

Prompt engineering requires a systematic approach centered on rigorous evaluations to track performance and identify regressions during model migrations. Maintaining production prompts involves applying general hygiene, such as using XML tags for structure and removing redundant, legacy instructions that cause overfitting. When models struggle with complex tasks, instructions alone are insufficient; providing external tools for calculations and implementing agentic workflows—such as generate-evaluate-repair loops—significantly improves reliability. These agentic systems allow for modular task handling, reducing token usage and latency while enabling the integration of dynamic, soft constraints at runtime. Ultimately, effective prompt maintenance relies on isolating failure modes and iteratively refining the system through data-driven testing rather than relying on broad, unverified instructions.