What problem does it solve? Iterating on AI skill prompts often relies on guesswork, leading to changes that feel better but actually regress quality, break evidence discipline, or overfit to a single test case. This Skill replaces ad-hoc prompt tweaking with a controlled experiment loop that produces measurable, documented improvements. ## Core Features & Use Cases - Controlled Mutation Loop: Establishes a baseline, changes one variable at a time, runs the same evaluation suite, and keeps or discards each mutation based on evidence. - Binary Eval Design Guidance: Includes a reference guide for writing yes/no evaluation criteria that resist gaming and produce reliable signals across runs. - Guardrails and Promotion Gate: Prevents optimizing toward fabricated claims or narrow metrics, and requires documented results before promoting a mutated skill to production. - Use Case: A marketing skill produces inconsistent positioning outputs. Define 4 binary evals, run a baseline, test targeted prompt mutations one at a time, log each experiment, and promote only the version that improves scores without regressions. ## Quick Start Use the autoresearch skill to run a controlled experiment improving the messaging-positioning skill with binary evals and a mutation log.
