rc-gan

# GAN Harness (generator ↔ evaluator) `make verify` proves a change is *correct*. It says nothing about whether a UI is well-designed, a CLI is pleasant, or copy lands. This skill closes that gap with an adversarial loop: a **generator** builds/improves the artifact, an independent **evaluator** exercises the *running* result and scores it against an explicit rubric, and the loop iterates on the evaluator's concrete feedback until the score clears a threshold or stops improving. The two roles are kept separate on purpose — nothing grades its own work, and the evaluator is fresh each round so it never anchors on the last verdict. ## Required Inputs - **Target**: what to build or improve (a feature, a screen, a command's UX, a document). - Optional **rubric**: dimensions + weights. If absent, derive one in Phase 0. - Optional **--threshold** (default `7.0` on a 1–10 scale) and **--max-iterations** (default `8`). ## Phase 0 — Spec & rubric Before iterating, write down what "good" means so scoring is not vibes: 1. Expand the target into a short spec: what it must do and the quality bar it must hit. 2. Define a **weighted rubric** of 3–5 dimensions summing to 1.0. Pick dimensions that fit the artifact, e.g. for a UI: `design 0.3 · craft 0.3 · functionality 0.2 · originality 0.2`; for a CLI: `ergonomics 0.4 · clarity 0.3 · correctness 0.2 · discoverability 0.1`. Each dimension scored 1–10; final = Σ(score × weight). 3. State how the artifact will be **exercised** (run it for ...