Gan Style Harness

This implements the adversarial generator-evaluator pattern from Anthropic's 2026 harness paper. Three agents work in sequence: a Planner expands your one-line prompt into a full product spec, a Generator builds it, and an Evaluator uses Playwright to test the live app and score it against design quality, originality, craft, and functionality. The Generator reads feedback and iterates until hitting a 7.0 weighted score or max iterations. Best for when you want production-quality output and can spend $50-200 on a full-stack app. Skip it for quick fixes or tight budgets. The key insight is that agents are terrible at evaluating their own work but you can engineer a separate evaluator to be ruthlessly strict.

skills.sh installs: 5.4K
GitHub stars: 300.3K
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add affaan-m/ecc --skill gan-style-harness --agent claude-code

Installs into .claude/skills of the current project.

Files

GAN-Style Harness Skill

Inspired by Anthropic's Harness Design for Long-Running Application Development (March 24, 2026)

A multi-agent harness that separates generation from evaluation, creating an adversarial feedback loop that drives quality far beyond what a single agent can achieve.

Core Insight

When asked to evaluate their own work, agents are pathological optimists — they praise mediocre output and talk themselves out of legitimate issues. But engineering a separate evaluator to be ruthlessly strict is far more tractable than teaching a generator to self-critique.

This is the same dynamic as GANs (Generative Adversarial Networks): the Generator produces, the Evaluator critiques, and that feedback drives the next iteration.