GAN-Style Harness
Build apps autonomously with a Generator-Evaluator harness: a planner writes the spec, a generator builds it, a ruthlessly strict evaluator tests the live app with Playwright, iterating until it clears the quality bar.
- What
- Build apps autonomously with a Generator-Evaluator harness: a planner writes the spec, a generator builds it, a ruthlessly strict evaluator tests the live app with Playwright, iterating until it clears the quality bar.
- Cost
- Free
- Needs
- a healthy API budget (expect roughly $50-200 per full run), Playwright MCP available for the evaluator, several hours of wall time.
- Level
- Beginner — a healthy API budget (expect roughly $50-200 per…
- Install
- Copy the installer prompt below into your Muse — your agent does the rest.
Curated by Skill Harbor: a GAN-inspired multi-agent harness that separates generation from evaluation, based on Anthropic's March 2026 harness design paper. The core insight is stated plainly: agents asked to evaluate their own work are pathological optimists, but engineering a separate evaluator to be ruthlessly strict is far more tractable than teaching a generator to self-critique. Three agents run the show: the Planner, a deliberately ambitious product manager that expands a one-line prompt into a multi-sprint spec with evaluation criteria; the Generator, a developer that negotiates a sprint contract with the Evaluator before writing code; and the Evaluator, a QA engineer that tests the live running application with Playwright, scoring design quality, originality, craft, and functionality on a weighted rubric with a 7.0 pass threshold, never praising mediocre work. Five to fifteen iterations later, you get the difference Anthropic published: 20 minutes and 9 dollars of solo-agent barely-functional output versus 4 to 6 hours and 125 to 200 dollars of production-ready quality. Configuration runs through environment variables for models, iterations, thresholds, criteria, and eval modes (playwright, screenshot, or code-only). Anti-patterns are documented: lenient evaluators, generators ignoring file-based feedback, infinite loops, superficial evaluator testing, evaluators praising their own fixes, and context exhaustion. The harness even encodes an evolution path: as models improve, strip away the scaffolding and re-test the assumptions. By @affaan-m, listed here with credit to its creator, inspired by Anthropic's paper by Prithvi Rajasekaran. From the affaan-m/ECC repository (MIT). Honest caveats: expect roughly 50 to 200 dollars per full run, so reserve this for projects where quality genuinely matters, never for quick fixes or tight budgets; the Evaluator needs Playwright MCP for real interaction testing. Skill Harbor never reviews the code, review it yourself before use.
Version:
Quick answer
GAN-Style Harness is a free Muse skill and GAN-inspired harness separating generation from evaluation, based on Anthropic's March 2026 paper. Its core insight is plain: agents evaluating their own work are optimists, but a separate strict evaluator is far more tractable. A planner writes the spec, a generator builds it, and the evaluator tests the live app with Playwright.
Install
Copy the install package below, then paste it into Muse🧪 Install ran in our test sandbox · Nov 11, 2026📌 Pinned version: commit 4eb71d9Pinned version: install the code at commit 4eb71d92a39cab44ad40ac9d8a6a5ccb4029d6c2 exactly (Skill Harbor tested this version). Do not substitute a newer version. If this commit is unreachable, stop and tell me. Use "GAN-Style Harness" with your Muse. Prerequisites: a healthy API budget (expect roughly $50-200 per full run), Playwright MCP available for the evaluator, several hours of wall time. 1. Open the skill: https://raw.githubusercontent.com/affaan-m/ECC/4eb71d92a39cab44ad40ac9d8a6a5ccb4029d6c2/skills/gan-style-harness/SKILL.md and copy the full SKILL.md text. 2. Paste it into a chat with Muse and add: "Plan the harness for this: [describe the app]. Keep the default 7.0 pass threshold." 3. Let the planner write the spec, review it, then start the generator-evaluator loop with /project:gan-build or the shell script. Tip: use gan-design (generator plus evaluator only) for frontend design work; keep the planner for full-stack builds. Safety: a skill is plain-text instructions; it runs nothing by itself. Review generated commands before running them.
Saved to your recent installs. Find it anytime on /connect.
🔒 Security, provenance and what this touches
The install prompt below already includes the vetting steps: your agent follows the community checklist before installing anything with executable code. Want more?
Verification certificate
Dated facts about what was checked for this listing, and how. This is not a safety certification.
- Pinned version
- 4eb71d9 · 2026-10-13
- Sandbox install test
- Install ran in our test sandbox · Nov 11, 2026
- Creator identity
- Not verified
- Fingerprint watch
- Last checked · Nov 21, 2026 : no change detected in the files linked by this install prompt
- Install prompt analysis
- 2026-10-13 : No commands, network calls, packages, file writes or API keys detected.
Skill Harbor never reviews the code of a listed skill. These are dated facts about the install prompt and our install test, not a safety certification. Review a skill yourself before you install it. How we verify
Questions
How do I install a skill?
Every product page includes a copy-paste install prompt. Paste it into your Muse and it sets the skill up for you — no manual configuration.
Where does my money go?
Straight to the seller. Skill Harbor never processes payments: checkout happens on the seller’s own page, usually Stripe.
What does the ✓ next to a creator’s name mean?
It means we confirmed the identity of the person behind the listing. It says nothing about the code itself — always check a skill before installing it.