claude-harness-builder-dev

Harness Builder — Claude Code Plugin

Your AI agent grades its own homework. It gets an A every time. The work is still broken.

This plugin fixes that.

What If Claude Could Actually QA Itself?

Every developer using AI hits the same wall: you ask it to build something, it says "Done!", and half the features are broken. You point out the bugs, it says "Fixed!", and introduces new ones. The agent literally cannot tell when its own work is mediocre.

Harness Builder uses the same GAN-inspired architecture from Anthropic's engineering team to split Claude into adversarial agents — one builds, one breaks. The builder never sees the evaluator's reasoning. The evaluator never sees the builder's excuses. Only hard evidence and scores decide if the code ships.

The result: production-quality apps from a 2-sentence prompt.

Install in 30 Seconds

Step 1: Add the marketplace

/plugin marketplace add https://github.com/vamgan/claude-harness-builder.git

Step 2: Install the plugin

/plugin install claude-harness-builder@claude-harness-builder-dev

That's it. Or load directly for testing:

claude --plugin-dir /path/to/claude-harness-builder

For your whole team — drop this in .claude/settings.json:

{
  "extraKnownMarketplaces": {
    "claude-harness-builder-dev": {
      "source": {
        "source": "url",
        "url": "https://github.com/vamgan/claude-harness-builder.git"
      }
    }
  },
  "enabledPlugins": {
    "claude-harness-builder@claude-harness-builder-dev": true
  }
}

How It Works

You type a prompt. Four agents do the rest:

flowchart TD
    A["Your idea\n'Build me a recipe app'"] --> B["Planner\nExpands into full product spec"]
    B --> C{You approve?}
    C -->|Yes| D["Contract Negotiator\nDefines testable success criteria"]
    C -->|Edit| B

    subgraph Sprint["Sprint Cycle (max 3)"]
        direction TB
        D --> E["Generator\nBuilds code, commits to git"]

        subgraph Loop["Adversarial Loop (max 5 iterations)"]
            direction TB
            E --> F["Evaluator\nTests independently, scores with evidence"]
            F --> G{Score >= 8/10?}
            G -->|Fail| H["Actionable feedback\n(no hand-waving)"]
            H --> E
        end

        G -->|Pass| I[Sprint Complete]
    end

    I --> J{More to build?}
    J -->|Yes| D
    J -->|No| K["Ship it"]

    style A fill:#e1f5fe
    style K fill:#c8e6c9
    style G fill:#fff9c4

The Before and After

Without Harness Builder	With Harness Builder
Agent builds and self-approves	Separate agents build and evaluate
"Looks good to me!" on broken code	Scores backed by actual command output
You manually QA everything	Independent evaluator catches what the builder missed
1 attempt, pray it works	Up to 15 automatic refinement cycles
Vague "I fixed it" responses	Sprint contracts with testable criteria before any code is written

The Secret: Information Firewall

flowchart LR
    subgraph G["Generator"]
        G1[Reads spec]
        G2[Reads contract]
        G3[Reads feedback.md]
        G4[Writes code]
    end

    subgraph E["Evaluator"]
        E1[Reads spec]
        E2[Reads contract]
        E3[Tests running app]
        E4[Writes feedback.md]
    end

    E4 -.->|"Only channel between agents"| G3

    style E4 fill:#c8e6c9
    style G3 fill:#c8e6c9

The Generator never sees the Evaluator's internal scores or reasoning. The Evaluator never sees the Generator's self-assessment. This prevents:

• Gaming — the builder can't learn to exploit the evaluator's scoring logic
• Anchoring — the evaluator isn't biased by the builder claiming "9/10"
• Rubber-stamping — the evaluator tests with real commands, not vibes

Sprint Contracts: No More "Trust Me, It Works"

Before a single line of code is written, agents negotiate testable success criteria:

#	Criterion	How It's Verified	Pass Threshold
1	Server starts without errors	npm start && curl localhost:3000	7/10
2	POST /api/recipes returns 201	curl -X POST -d '{"name":"test"}'	7/10
3	Search filters by ingredient	curl /api/recipes?ingredient=tomato	7/10
4	Invalid input returns 400	curl -X POST -d '{}'	7/10

Pass conditions: Average >= 8/10. No criterion below 7/10. Any criterion below 5 = automatic fail.

No subjective hand-waving. No "the code looks clean." Every criterion is verified with an actual command.

Usage

Just tell Claude what to build:

Build me a recipe sharing app where users can post recipes and search by ingredient

The harness activates automatically. Or invoke directly: /claude-harness-builder:harness-builder