What problem does it solve? One-shot generation often misses the bar when a task has a concrete, checkable standard — a failing test suite, a lint rubric, or stated document requirements. This Skill turns that into a controlled convergence loop: generate a candidate, evaluate it against explicit criteria, feed the specific failure back into the next attempt, and stop when it passes or a round cap is hit. ## Core Features & Use Cases - Objective-first evaluation: Runs real checks (test suites, go vet, swiftlint, builds) as ground truth before falling back to model-as-evaluator for criteria that cannot be mechanically checked. - Separated generate/evaluate/refine roles: The evaluator grades against stated criteria without silently rewriting, keeping feedback a clean signal for the refinement step. - Capped convergence with escalation: Stops after 3-5 rounds and surfaces persistent failure as a signal that criteria, specification, or approach is wrong rather than spinning indefinitely. - Use Case: A bug fix must pass a test suite — generate the fix, run the tests, feed the exact failure output into the next attempt, and repeat until green or the cap triggers a rethink. ## Quick Start Ask the AI to fix the failing test by iterating with the evaluator-optimizer loop: generate a fix, run the test suite, and refine from the failure output until it passes or four rounds elapse.
