Skip to content
Mark Warren
Menu

All work

Coding agents workflow

Running Claude Code and Codex on the same repositories

How I work with Claude Code and Codex together: one instruction source, custom skills written once, a second model that critiques every plan, and rules that keep the agents from destroying each other's work.

How it fits together

  1. Inputs

    • One instruction source

      Global rules live in one file and are installed for both tools, with a check that reports drift.

    • Repository canon

      Each repository's AGENTS.md is canon, and its CLAUDE.md only imports it.

  2. Engine

    Two agents

    Claude Code and Codex work the same repositories under the same rules.

    • Claude
    • Codex
    1. Custom skills

      Discovery prep, call briefs, proposal drafts, delivery status, idea review, and plan review, written once and mirrored for the second tool.

    2. Plan, critique, decide

      Claude drafts a plan, Codex critiques it in a read-only sandbox, and every finding is accepted or rejected in writing before building.

      • Codex
    3. One agent per tree

      Parallel work goes in separate worktrees on separate branches.

  3. Outputs

    • Reviewed plans

      A plan, its critique, and a final plan with dispositions, kept with the code.

    • Repeatable work

      Recurring tasks run the same written method each time, in either tool.

Guardrails

  • Pushes, deploys, production writes, sends, and spend wait for a person
  • Uncommitted work is never discarded
  • Plan review runs on a subscription, with no metered spend

The problem

Two coding agents on the same repositories can drift apart in their instructions and can overwrite each other's uncommitted work. A single model reviewing its own plan also tends to agree with itself.

What I built

  • One source of global instructions, installed for Claude Code and Codex by script, with a drift check.
  • A repository convention where AGENTS.md is canon for every tool and skills are authored in one folder and mirrored to the other.
  • Custom skills for the recurring work of a services company, from discovery prep to proposal drafts to delivery status.
  • A dual-model plan review that runs Codex read-only against the repository and records a disposition for every finding.
  • Concurrency rules: one agent per working tree, worktrees for parallel work, and a hard stop when a tree holds changes the agent did not make.

How I built it

The rules are plain Markdown under version control, the installers and mirrors are small PowerShell scripts, and the plan review calls the Codex command line in a read-only sandbox with the user's own configuration ignored. This page was itself planned that way: Codex critiqued the plan before any copy was written.

What it shows

I use AI coding tools as a managed team with written rules, review, and limits, not as an autocomplete.