What problem does it solve? Prompt authors cannot objectively judge their own instructions, so ambiguous wording in skills, slash commands, and task prompts goes undetected until agents misbehave. This Skill replaces guesswork with empirical evaluation: unbiased subagents execute the prompt, and both self-reported feedback and measured metrics drive iterative refinement. ## Core Features & Use Cases - Bias-Free Execution: Dispatches fresh subagents per iteration to run evaluation scenarios against the target prompt, avoiding author bias. - Dual-Sided Metrics: Combines executor self-reports (ambiguities, discretion gaps, retries) with measured signals (success rate, accuracy, tool_uses, duration) against a fixed requirements checklist. - Convergence & Overfitting Control: Stops after consecutive clean iterations, runs hold-out scenarios to detect overfitting, and flags divergence when the prompt design itself is flawed. - Use Case: After writing a new skill or heavily revising a CLAUDE.md section, run 2-3 realistic scenarios through fresh subagents, record ambiguities and accuracy per iteration, and apply one minimal fix per cycle until metrics plateau. ## Quick Start Evaluate my newly written skill prompt by running it through fresh subagents on two realistic scenarios and iterate on the ambiguities they report.
