eval-fix

Iteratively generate and evaluate fixes until feature gaps converge.

8|1|Updated Mar 22, 2026
One-click install
npx skills add https://github.com/zxdxjtu/harness --skill eval-fix
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: eval-fix
Source: https://github.com/zxdxjtu/harness/tree/main/skills/eval-fix
Command: npx skills add https://github.com/zxdxjtu/harness --skill eval-fix

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This skill implements an adversarial fix-evaluate loop that separates generation of fixes from evaluation, enabling systematic, iterative repair of feature gaps until convergence.

Core Features & Use Cases

  • Adversarial fix generation: Generator proposes concrete fixes based on eval findings and reference evidence.
  • Independent evaluation: Evaluator validates fixes against evaluation criteria and screenshots to prevent self-assessment bias.
  • Iterative convergence: Repeats fix and evaluation cycles until a target score or manual intervention is reached.
  • Progress tracking: Logs iterations, scores, and remaining gaps to guide next steps.

Quick Start

Run /evaluate for a feature to generate an eval report, then start the fix loop with /eval-fix to iteratively generate and verify fixes.

Frequently Asked Questions about eval-fix

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How does an adversarial fix loop improve quality assurance for feature implementations?

An adversarial fix loop improves quality assurance by strictly separating fix generation from evaluation, forcing an independent evaluator to validate proposed fixes against criteria and evidence, iterating until convergence to prevent self-assessment bias.

How do I start an iterative fix and evaluate cycle to repair feature gaps?

To start an iterative fix and evaluate cycle, first generate an evaluation report for the feature, then launch the fix loop to automatically propose concrete fixes, validate them independently, and log iteration progress and convergence metrics.

Can I use an automated fix loop to converge feature implementations without manual intervention?

Yes, you can use an automated fix loop to converge feature implementations without manual intervention by setting a target score; the loop repeats generation and evaluation cycles automatically until that score is reached or you manually stop it.

What is the best way to prevent self-assessment bias during automated feature repair?

The best way to prevent self-assessment bias during automated feature repair is enforcing strict role separation where an independent evaluator validates generator fixes against evaluation reports and screenshots, ensuring objective convergence tracking.

Why does my automated fix loop require evidence from evaluation reports?

An automated fix loop requires evidence from evaluation reports to ensure the generator proposes concrete fixes based on actual feature gaps and the evaluator can objectively validate improvements, enabling systematic iterative convergence.

When should I manually intervene in an automated fix and evaluate workflow?

You should manually intervene in an automated fix and evaluate workflow when the iteration fails to converge to the target score, leaving remaining gaps that require human guidance to progress beyond the automated loop.