prompt-engineer-agent-prompts-evals

Offline eval suites for tool-using agents require golden datasets, judge rubrics, and regression CI checks to validate behavior. This Skill provides structured guidance for designing evaluation pipelines, offline eval harnesses, and release gates to ensure reliable agent performance.

Prompt governance defines team roles, change control, and release gates for modifying system and developer prompts. This Skill documents best practices for coordinating cross-team prompt deployment, ensuring prompt changes pass evaluation pipelines and policy requirements before release.

Regression CI checks for agent prompt changes integrate offline eval harnesses with golden datasets to catch behavioral regressions. This Skill guides you through designing evaluation pipelines, defining success metrics, and implementing release gates to validate prompt modifications.

Structuring system prompts, developer prompts, and tool schemas for agents requires robust prompt engineering and clear handoff definitions. This Skill identifies best practices for prompt design, tool schema configuration, and agent handoffs to achieve reliable tool-using behavior.

Judge rubrics enable automated evaluation of agent behavior within offline eval pipelines. This Skill provides guidance for developing judge rubrics, defining success metrics, and integrating them into evaluation pipelines to govern prompt changes and ensure agent reliability.

Golden datasets are essential for testing agent prompt modifications through offline eval harnesses and regression CI. This Skill covers designing golden datasets, interfacing with eval datasets, and establishing release gates to validate that prompt changes maintain expected behavior.