What problem does it solve? Choosing a Claude effort level (low through max) is often guesswork carried over from older models, leading to slow or costly runs with no measured quality gain. This Skill resolves the level currently in effect, proposes a starting level from task shape, and optionally runs a controlled A/B eval to pick the lowest level that matches the best quality. ## Core Features & Use Cases - Effective Level Resolution: Walks the precedence chain from environment variable and CLI flag through frontmatter and settings.json to report the level in effect and its source. - Starting Level Recommendation: Maps task shape and model (Opus 5.5, Fable 5.1, Sonnet 5) to a starting effort level and names exactly where to set it per scope. - Optional A/B Eval: Runs a frozen prompt at two or three adjacent levels via claude -p, scores against a rubric fixed before the runs, and picks the lowest level tying the best median score. - Use Case: A scheduled claude -p summary job runs slowly after a model upgrade. Use this Skill to resolve its current level, run a six-run eval across low and medium, and record that low ties on quality at half the cost. ## Quick Start Ask the AI to recommend an effort level for your scheduled claude -p job and, if you approve the run count and cost, run a small A/B eval to confirm the cheapest level that holds quality.
