gan-style-harness

# GAN-Style Harness Skill

> Inspired by [Anthropic's Harness Design for Long-Running Application Development](https://www.anthropic.com/engineering/harness-design-long-running-apps) (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.

## When to Use

- Building complete applications from a one-line prompt
- Frontend design tasks requiring high visual quality
- Full-stack projects that need working features, not just code
- Any task where "AI slop" aesthetics are unacceptable
- Projects where you want to invest $50-200 for production-quality output

## When NOT to Use

- Quick single-file fixes (use standard `claude -p`)
- Tasks with tight budget constraints (<$10)
- Simple refactoring (use de-sloppify pattern instead)
- Tasks that are already well-specified with tests (use TDD workflow)

## Architecture

```
                    ┌─────────────┐
                    │   PLANNER   │
                    │  (Opus 4.6) │
                    └──────┬──────┘
                           │ Product Spec
                           │ (features, sprints, design direction)
                           ▼
              ┌────────────────────────┐
              │                        │
              │   GENERATOR-EVALUATOR  │
              │      FEEDBACK LOOP     │
              │                        │
              │  ┌──────────┐          │
              │  │GENERATOR │--build-->│──┐
              │  │(Opus 4.6)│          │  │
              │  └────▲─────┘          │  │
              │       │                │  │ live app
              │    feedback             │  │
              │       │                │  │
              │  ┌────┴─────┐          │  │
              │  │EVALUATOR │<-test----│──┘
              │  │(Opus 4.6)│          │
              │  │+Playwright│         │
              │  └──────────┘          │
              │                        │
              │   5-15 iterations      │
              └────────────────────────┘
```

## The Three Agents

### 1. Planner Agent

**Role:** Product manager — expands a brief prompt into a full product specification.

**Key behaviors:**
- Takes a one-line prompt and produces a 16-feature, multi-sprint specification
- Defines user stories, technical requirements, and visual design direction
- Is deliberately **ambitious** — conservative planning leads to underwhelming results
- Produces evaluation criteria that the Evaluator will use later

**Model:** Opus 4.6 (needs deep reasoning for spec expansion)

### 2. Generator Agent

**Role:** Developer — implements features according to the spec.

**Key behaviors:**
- Works in structured sprints (or continuous mode with newer models)
- Negotiates a "sprint contract" with the Evaluator before writing code
- Uses full-stack tooling: React, FastAPI/Express, databases, CSS
- Manages git for version control between iterations
- Reads Evaluator feedback and incorporates it in next iteration

**Model:** Opus 4.6 (needs strong coding capability)

### 3. Evaluator Agent

**Role:** QA engineer — tests the live running application, not just code.

**Key behaviors:**
- Uses **Playwright MCP** to interact with the live application
- Clicks through features, fills forms, tests API endpoints
- Scores against four criteria (configurable):
  1. **Design Quality** — Does it feel like a coherent whole?
  2. **Originality** — Custom decisions vs. template/AI patterns?
  3. **Craft** — Typography, spacing, animations, micro-interactions?
  4. **Functionality** — Do all features actually work?
- Returns structured feedback with scores and specific issues
- Is engineered to be **ruthlessly strict** — never praises mediocre work

**Model:** Opus 4.6 (needs strong judgment + tool use)

## Evaluation Criteria

The default four criteria, each scored 1-10:

```markdown
## Evaluation Rubric

### Design Quality (weight: 0.3)
- 1-3: Generic, template-like, "AI slop" aesthetics
- 4-6: Competent but unremarkable, follows conventions
- 7-8: Distinctive, cohesive visual identity
- 9-10: Could pass for a professional designer's work

### Originality (weight: 0.2)
- 1-3: Default colors, stock layouts, no personality
- 4-6: Some custom choices, mostly standard patterns
- 7-8: Clear creative vision, unique approach
- 9-10: Surprising, delightful, genuinely novel

### Craft (weight: 0.3)
- 1-3: Broken layouts, missing states, no animations
- 4-6: Works but feels rough, inconsistent spacing
- 7-8: Polished, smooth transitions, responsive
- 9-10: Pixel-perfect, delightful micro-interactions

### Functionality (weight: 0.2)
- 1-3: Core features broken or missing
- 4-6: Happy path works, edge cases fail
- 7-8: All features work, good error handling
- 9-10: Bulletproof, handles every edge case
```

### Scoring

- **Weighted score** = sum of (criterion_score * weight)
- **Pass threshold** = 7.0 (configurable)
- **Max iterations** = 15 (configurable, typically 5-15 sufficient)

## Usage

### Via Command

```bash
# Full three-agent harness
/project:gan-build "Build a project management app with Kanban boards, team collaboration, and dark mode"

# With custom config
/project:gan-build "Build a recipe sharing platform" --max-iterations 10 --pass-threshold 7.5

# Frontend design mode (generator + evaluator only, no planner)
/project:gan-design "Create a landing page for a crypto portfolio tracker"
```

### Via Shell Script

```bash
# Basic usage
./scripts/gan-harness.sh "Build a music streaming dashboard"

# With options
GAN_MAX_ITERATIONS=10 \
GAN_PASS_THRESHOLD=7.5 \
GAN_EVAL_CRITERIA="functionality,performance,security" \
./scripts/gan-harness.sh "Build a REST API for task management"
```

### Via Claude Code (Manual)

```bash
# Step 1: Plan
claude -p --model opus "You are a Product Planner. Read PLANNER_PROMPT.md. Expand this brief into a full product spec: 'Build a Kanban board app'. Write spec to spec.md"

# Step 2: Generate (iteration 1)
claude -p --model opus "You are a Generator. Read spec.md. Implement Sprint 1. Start the dev server on port 3000."

# Step 3: Evaluate (iteration 1)
claude -p --model opus --allowedTools "Read,Bash,mcp__playwright__*" "You are an Evaluator. Read EVALUATOR_PROMPT.md. Test the live app at http://localhost:3000. Score against the rubric. Write feedback to feedback-001.md"

# Step 4: Generate (iteration 2 — reads feedback)
claude -p --model opus "You are a Generator. Read spec.md and feedback-001.md. Address all issues. Improve the scores."

# Repeat steps 3-4 until pass threshold met
```

## Evolution Across Model Capabilities

The harness should simplify as models improve. Following Anthropic's evolution:

### Stage 1 — Weaker Models (Sonnet-class)
- Full sprint decomposition required
- Context resets between sprints (avoid context anxiety)
- 2-agent minimum: Initializer + Coding Agent
- Heavy scaffolding compensates for model limitations

### Stage 2 — Capable Models (Opus 4.5-class)
- Full 3-agent harness: Planner + Generator + Evaluator
- Sprint contracts before each implementation phase
- 10-sprint decomposition for complex apps
- Context resets still useful but less critical

### Stage 3 — Frontier Models (Opus 4.6-class)
- Simplified harness: single planning pass, continuous generation
- Evaluation reduced to single end-pass (model is smarter)
- No sprint structure needed
- Automatic compaction handles context growth

> **Key principle:** Every harness component encodes an assumption about what the model can't do alone. When models improve, re-test those assumptions. Strip away what's no longer needed.

## Configuration

### Environment Variables

| Variable | Default | Description |
|----------|---------|-------------|
| `GAN_MAX_ITERATIONS` | `15` | Maximum generator-evaluator cycles |
| `GAN_PASS_THRESHOLD` | `7.0` | Weighted score to pass (1-10) |
| `GAN_PLANNER_MODEL` | `opus` | Model for planning agent |
| `GAN_GENERATOR_MODEL` | `opus` | Model for generator agent |
| `GAN_EVALUATOR_MODEL` | `opus` | Model for evaluator agent |
| `GAN_EVAL_CRITERIA` | `design,originality,craft,functionality` | Comma-separated criteria |
| `GAN_DEV_SERVER_PORT` | `3000` | Port for the live app |
| `GAN_DEV_SERVER_CMD` | `npm run dev` | Command to start dev server |
| `GAN_PROJECT_DIR` | `.` | Project working directory |
| `GAN_SKIP_PLANNER` | `false` | Skip planner, use spec directly |
| `GAN_EVAL_MODE` | `playwright` | `playwright`, `screenshot`, or `code-only` |

### Evaluation Modes

| Mode | Tools | Best For |
|------|-------|----------|
| `playwright` | Browser MCP + live interaction | Full-stack apps with UI |
| `screenshot` | Screenshot + visual analysis | Static sites, design-only |
| `code-only` | Tests + linting + build | APIs, libraries, CLI tools |

## Anti-Patterns

1. **Evaluator too lenient** — If the evaluator passes everything on iteration 1, your rubric is too generous. Tighten scoring criteria and add explicit penalties for common AI patterns.

2. **Generator ignoring feedback** — Ensure feedback is passed as a file, not inline. The generator should read `feedback-NNN.md` at the start of each iteration.

3. **Infinite loops** — Always set `GAN_MAX_ITERATIONS`. If the generator can't improve past a score plateau after 3 iterations, stop and flag for human review.

4. **Evaluator testing superficially** — The evaluator must use Playwright to **interact** with the live app, not just screenshot it. Click buttons, fill forms, test error states.

5. **Evaluator praising its own fixes** — Never let the evaluator suggest fixes and then evaluate those fixes. The evaluator only critiques; the generator fixes.

6. **Context exhaustion** — For long sessions, use Claude Agent SDK's automatic compaction or reset context between major phases.

## Results: What to Expect

Based on Anthropic's published results:

| Metric | Solo Agent | GAN Harness | Improvement |
|--------|-----------|-------------|-------------|
| Time | 20 min | 4-6 hours | 12-18x longer |
| Cost | $9 | $125-200 | 14-22x more |
| Quality | Barely functional | Production-ready | Phase change |
| Core features | Broken | All working | N/A |
| Design | Generic AI slop | Distinctive, polished | N/A |

**The tradeoff is clear:** ~20x more time and cost for a qualitative leap in output quality. This is for projects where quality matters.

## References

- [Anthropic: Harness Design for Long-Running Apps](https://www.anthropic.com/engineering/harness-design-long-running-apps) — Original paper by Prithvi Rajasekaran
- [Epsilla: The GAN-Style Agent Loop](https://www.epsilla.com/blogs/anthropic-harness-engineering-multi-agent-gan-architecture) — Architecture deconstruction
- [Martin Fowler: Harness Engineering](https://martinfowler.com/articles/exploring-gen-ai/harness-engineering.html) — Broader industry context
- [OpenAI: Harness Engineering](https://openai.com/index/harness-engineering/) — OpenAI's parallel work