Codex Harness Public

A layered Codex harness for long-term knowledge work, reusable workflows, project bootstrapping, and project-specific AGENTS generation.

This repository packages a public version of my Codex harness system. It is designed to make Codex behave less like a one-off chat assistant and more like a reusable working system for multi-project collaboration.

What This Repository Includes

• AGENTS.md Global collaboration rules, task routing principles, delivery standards, and reliability boundaries.
• knowledge-workbench/ A modular workbench for recurring workflows, including templates, playbooks, style guidance, quality gates, scoring, calibration, and benchmark routines.
• skills/project-harness-bootstrap/ A reusable skill that turns a project description into project-root harness files such as AGENTS.md and PROJECT-WORKFLOW.md.

What This Harness Tries To Do

• Keep the global layer small and stable.
• Push detailed workflows down into reusable modules instead of bloating top-level instructions.
• Help Codex ask fewer vague questions and move work forward more directly.
• Preserve output quality for writing, review, presentation, and structured knowledge work.
• Make it easier to start a new project with a usable project-level harness.

Architecture

This harness uses a layered structure:

1. Global layer AGENTS.md holds stable cross-project rules such as collaboration style, task routing, delivery expectations, and reliability constraints.

2. Workbench layer knowledge-workbench/ holds reusable building blocks:

   • templates/ for common output structures
   • playbooks/ for routing, delivery, quality, and failure recovery
   • style/ for aesthetic direction and style translation
   • architecture/ for rule precedence, scoring, and harness governance
   • benchmarks/ for calibration and regression checks

3. Project layer Each project can define its own root AGENTS.md and PROJECT-WORKFLOW.md based on its actual purpose, outputs, and recurring workflows.

Who This Is For

This repository is most useful if you use Codex for:

• knowledge work
• writing and revision
• research and review
• presentation and deck preparation
• multi-project workflow setup
• long-term agent collaboration

Installation

Copy the files into your Codex home directory.

Recommended layout:

~/.codex/
  AGENTS.md
  knowledge-workbench/
  skills/
    project-harness-bootstrap/

Install in this order:

1. Copy AGENTS.md into ~/.codex/AGENTS.md
2. Copy knowledge-workbench/ into ~/.codex/knowledge-workbench/
3. Copy skills/project-harness-bootstrap/ into ~/.codex/skills/project-harness-bootstrap/

If you already have your own global AGENTS.md, merge carefully instead of overwriting blindly.

Quick Start

After installation, the global harness will shape Codex behavior across projects.

For a new project, use the bootstrap skill with a short project description.

Example:

Use $project-harness-bootstrap to create a project-root harness from this project description.

This is a research and writing project about...
The recurring tasks are...
The main deliverables are...
The preferred tools are...
Actions that require confirmation are...

Or in Chinese:

用 $project-harness-bootstrap 根据下面这段项目描述，在项目根目录生成 AGENTS.md 和 PROJECT-WORKFLOW.md。

这个项目是做……
核心重复任务有……
主要交付物有……
希望优先使用的工具有……
需要确认的边界有……

What The Bootstrap Skill Generates

By default, the skill creates or updates:

• AGENTS.md A project-specific Codex harness that changes behavior inside that project.
• PROJECT-WORKFLOW.md A human-readable quickstart and workflow map for the project.

The skill is designed to:

• inherit your global rules
• avoid duplicating the global harness
• keep project instructions compact
• turn project descriptions into usable workflows quickly

Repository Structure

codex-harness-public/
  AGENTS.md
  README.md
  knowledge-workbench/
    architecture/
    benchmarks/
    playbooks/
    questioning/
    style/
    templates/
  skills/
    project-harness-bootstrap/
      SKILL.md
      agents/
      references/

Design Principles

• Keep the global layer small.
• Keep project workflows project-specific.
• Prefer routing over repetition.
• Prefer reusable playbooks over long top-level prompts.
• Prefer explicit delivery standards over vague quality language.
• Prefer repair loops and benchmarks over endlessly adding more rules.

What Is Not Included

This public package intentionally excludes private local state such as:

• personal config files
• auth files
• plugin caches
• session data
• local databases
• API keys
• machine-specific automations

Suggested Next Steps

If you want to publish this as your own repository, consider adding:

• consider adding examples / changelog
• a short CHANGELOG.md
• example project descriptions
• example generated project-level harness outputs