What problem does it solve? Prompt authors cannot objectively judge their own instructions, so ambiguous wording in skills, slash commands, CLAUDE.md files, and code-generation prompts causes agents to misbehave in ways the author never anticipated. This Skill replaces guesswork with empirical testing: a fresh subagent actually executes the prompt, and both sides of the exchange are measured until improvements plateau. ## Core Features & Use Cases - Bias-free execution testing: Dispatches a blank-slate subagent to run the target prompt against 2-3 realistic scenarios, then collects structured self-reports on unclear points, discretionary fill-ins, and retries. - Quantified iteration loop: Tracks success/failure, accuracy against a fixed requirements checklist, step count, duration, and retry count per iteration, with explicit convergence, divergence, and overfitting criteria. - Failure pattern ledger: Accumulates class-level fix rules across iterations so recurring failure modes are detected instead of re-patched, plus an optional two-variant exploration mode for breaking plateaus. - Use Case: After writing a new skill whose description promises navigation, form filling, and data extraction, run this Skill to verify the body actually delivers, fix the gaps iteration by iteration, and stop only when two consecutive rounds show zero new unclear points. ## Quick Start Empirically evaluate and iteratively improve the prompt in my skill file by dispatching fresh subagents against realistic scenarios until the improvements converge.
