PolyArch/humanize

What it solves

Humanize is a plugin for Claude Code that replaces the "one-shot" approach to AI coding with an iterative development cycle. It prevents blind spots in AI-generated code by introducing an independent review process, ensuring that code is refined incrementally until it meets all acceptance criteria.

How it works

The project implements a workflow called RLCR (Ralph-Loop with Codex Review). This process splits development into two distinct phases:

1. Implementation: Claude handles the actual coding tasks.
2. Code Review: An independent agent (Codex) reviews the code quality and marks issues with severity markers.

These issues are then fed back into the implementation phase, creating a continuous loop of refinement. The system also includes tools for generating initial idea drafts and detailed plans before the loop begins, ensuring the human user remains the architect of the project.

Who it’s for

Developers using Claude Code who want a more rigorous, review-driven development process to reduce bugs and increase confidence in AI-generated code.

Highlights

• Independent Review: Uses a separate AI (Codex) to review Claude's work, eliminating single-model blind spots.
• Iterative Refinement: Continues the development loop until all acceptance criteria are passed.
• Swarm Mode: Ability to parallelize refinement using Agent Teams.
• Planning Tools: Built-in commands to generate and refine idea drafts and implementation plans.
• External Research: Integration with Gemini CLI for deep web research during development.