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 real): a web UI via a browser/dev server, a CLI by running commands, a doc by reading it against its goal. Evaluation is of the running thing, never of the source alone.

The loop

Repeat until stop condition. Track score per iteration.

1. Generate — implement or improve the artifact to satisfy the spec and address the previous iteration's feedback. Keep it real and runnable; leave the artifact in a runnable state. Run the project's correctness gate (make verify or equivalent) so quality work never ships a broken build.
2. Evaluate (fresh & independent) — spawn a separate reviewer (a subagent / fresh context — do not reuse the generator's reasoning). It must:
   • Actually exercise the running artifact against the "how to exercise" plan from Phase 0.
   • Score each rubric dimension 1–10 with one concrete justification each, then compute the weighted total.
   • Write specific, addressable defects ("the primary button has no hover state", "--help doesn't list the sync subcommand"), not vague notes. Penalize generic AI-slop (default gradients, lorem-ipsum, boilerplate layouts).
3. Decide:
   • Pass — weighted score ≥ threshold → stop, report success.
   • Plateau — score improved by < 0.3 for 2 consecutive iterations → stop; report the plateau and the best version (more iterations won't help; escalate to the user).
   • Cap — reached --max-iterations → stop; report the best version and remaining gaps.
   • Otherwise feed the evaluator's defects into the next Generate.