empirical-prompt-tuning

To evaluate skill definitions without bias, use a multi-step process that separates baseline evaluation from improvement proposals, ensuring changes are traceable to initial weaknesses before re-evaluation and convergence checks.

A convergence check in prompt tuning verifies that iterative improvements to skill definitions have stabilized, ensuring that further modifications no longer yield significant changes across evaluation rounds.

You assess the clarity and feasibility of a SKILL.md file by running a baseline evaluation across five axes: clarity, completeness, feasibility, quality guards, and integration, generating structured reports on outcomes.

Yes, you can automate iterative refinement by applying a staged workflow that generates improvement proposals, re-evaluates the results, and performs formal convergence checks across multiple rounds.

The best way to ensure governance before deployment is to apply a formal convergence protocol with calibration references, evaluating skill definitions for completeness and quality guards across structured reporting rounds.

Skill definitions fail to converge during prompt tuning when improvement proposals lack traceability to baseline weaknesses, preventing stable re-evaluation outcomes across iterative rounds.