What problem does it solve? Prompt authors cannot objectively judge their own instructions—what seems clear to the writer often confuses a fresh executor. This Skill removes author bias by dispatching fresh subagents to actually execute the prompt, then measuring results from both sides (executor self-reports plus instruction-side metrics) and iterating until improvements plateau. ## Core Features & Use Cases - Bias-Free Execution Testing: Dispatches new subagents per iteration to run evaluation scenarios against the target prompt, never reusing agents or relying on self-review. - Dual-Sided Metrics: Combines executor self-reports (ambiguities, discretion-filled gaps, retries) with measured metrics (success/failure, accuracy %, tool_uses, duration_ms) tracked in a fixed requirements checklist with [critical] tags. - Convergence & Overfitting Guards: Stops after 2 consecutive clean iterations with metric thresholds, and validates against a hold-out scenario to detect overfitting. - Use Case: After writing a new skill or slash command, run 2-3 realistic scenarios through fresh subagents, collect ambiguity reports, apply one minimal fix per iteration, and repeat until no new issues surface. ## Quick Start Use the empirical-prompt-tuning skill to evaluate and iteratively improve my newly written skill prompt with fresh subagent test runs.
