Review a target — code, a script, a file, a process, a workflow, a prompt, a config, a plan — for **efficiency**, and hand back **prioritized, location-anchored recommendations**, each carrying a per-candidate **no-regression verdict**. This skill **recommends**; it does not rewrite, and it does not edit the target. The shell tools it holds (`Bash`, `PowerShell`) are for **measuring and inspecting only** — running a benchmark, reading a file, counting tokens — **never for modifying the target.** (Tooling can't enforce that with a shell present; it's a discipline, stated here so the boundary is explicit.) To execute the recommendations, pass them to `/improve`. Natural pip…
