Hooligan harness

Implementing a Generator-Evaluator Architecture for AI Harnesses as a SKILL in Claude Code

Install

npx -y skills add aditikilledar/hooligan-harness

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Implements a high-reliability "Harness Engineering" loop using Planner, Generator, and Evaluator personas. Trigger when a user wants to "implement a feature," "start the harness," or "build with verification."

SKILL.md

2.5 KB, as published. Nobody here has run it

Objective

To replace one-shot code generation with a structured, self-correcting agentic loop that ensures all code is planned in YAML, implemented via best practices of coding, and verified by an adversarial evaluator before declared complete.

Instructions

1. Phase 0: Initialization

Read the repository README.md to understand the environment.
Create .harness/dev_init.md with instructions to run the development server for downstream agents.

2. Phase 1: Planning (The Planner)

Create a technical roadmap at .harness/[nickname].yaml.
Define specific, quantifiable Acceptance Criteria (AC) for every task.
Initialize an append-only log at .harness/progress.md to track all session activity.

3. Phase 2: Implementation (The Generator)

Select Task: Identify the next pending task based on depends_on logic.
Logic Synthesis: Perform an impact analysis and define a testing strategy before writing code.
Code Generation: Implement logic following SOLID, DRY, and KISS principles.
Atomic Updates: Every task completion requires a git commit and a progress entry.

4. Phase 3: Adversarial Evaluation (The Evaluator)

Hostile Barrier: Assume the Generator's output is riddled with bugs and happy-path logic.
Instant Death Gates: Immediately FAIL the task if there is a global regression, linting error, or any lazy code such as placeholders like TODO or FIXME.
AC Deep-Dive: Confirm every AC has a dedicated test and that test quality metrics like Mock Integrity and Coverage Stability are met.
Verdict: Return a binary PASS or FAIL. If FAIL, provide a root cause analysis.

5. Phase 4: Remediation and Reconciliation

Remediation: If the Evaluator returns FAIL, the Generator must suspend new work, reproduce the failure locally, and fix the logic until it passes evaluation.
Reconciliation: Once PASS is achieved, update the YAML task status to done and log the verification evidence including the git hash and test results in progress.md.

Reference

Persona - Planner: Focuses on structured YAML roadmap creation and task decomposition.
Persona - Generator: Focuses on defensive programming, architectural synthesis, and local verification.
Persona - Evaluator: Acts as the gatekeeper using a Zero-Trust approach to code quality.