empirical-prompt-tuning

Iterative prompt improvement requires running target prompts through subagents and collecting dual feedback to evaluate and refine instructions until performance stabilizes. This empirical prompt tuning process uses structured metrics to eliminate ambiguity and validate behavior.

Empirical prompt tuning evaluates agent instructions by combining the subagent's self-reported execution feedback with the instruction-side evaluation metrics. This dual-sided approach identifies behavioral mismatches and drives iterative refinement cycles until improvements plateau.

Validate revised agent skills or slash commands by dispatching subagents to execute the target instructions and report back with structured evaluation metrics. Repeatable cycles driven by YAML-frontmatter configuration ensure auditable validation with tool usage and duration reports.

Empirical prompt evaluation is needed immediately after creating or substantially revising agent prompts, skills, or slash commands. It stabilizes unpredictable agent behavior by running repeatable evaluation cycles with audit-ready reports containing tool usage history and execution duration.

Empirical prompt tuning operates without external dependencies or component libraries. It relies entirely on dispatching internal subagents and utilizing YAML-frontmatter driven configuration to structure the evaluation metrics and generate audit-ready reports.

Subagent feedback for prompt evaluation is limited by the plateauing of iterative improvements over repeated cycles. While dual-sided metrics reduce ambiguity, the process requires running until improvements stall, which may consume significant execution time and tool usage overhead.