“I created the code in Claude, had it do a code review, everything was fine. Then I pointed Codex at the same folder and said do a code review — and didn’t it find another half-dozen things Claude hadn’t found.” — Lou
Session context: 2026-05-28_Mastermind — Donald described shipping an app for his book using Claude/Opus to plan and Codex to execute; Lou shared his own parallel discovery, concluding he may finally open a second (OpenAI) subscription alongside Claude.
Core Idea
The instinct is to do all your building inside one model. Both Donald and Lou independently arrived at a better pattern: split the work across models by their comparative strengths. Opus (Claude) plans — it’s where the spec, the brainstorming, the architecture, and the knowledge work happen. Codex executes — once the spec is detailed enough, Codex implements it, and it does so with much larger usage limits and strong execution discipline. Donald’s loop was concrete: Opus planned, Codex executed against the plan and reported back at update gates, then Claude checked the result.
The second, sharper benefit is cross-model review. Lou had Claude write code, audit it, write tests — everything passed. Then he pointed Codex at the same folder and asked for a code review, and it surfaced half a dozen issues Claude had missed. A single model has blind spots in its own work; a second model with a different training lineage catches what the first cannot see. This is multi-model quality control applied to code specifically, and it’s only practical because the work lives in a folder on disk — which lets you bring different AIs to the same context without re-uploading anything. Files-on-disk isn’t just durable memory; it’s what makes a multi-vendor workflow frictionless.
Lou’s tentative economic conclusion: pair one Claude subscription with one ChatGPT/Codex subscription. Use Opus for spec and thinking; if the spec is detailed enough, Codex can “go away and implement the heck out of it” — and you get far more implementation runway out of Codex’s limits than out of Claude’s.
Practical Application
If you build with AI and keep hitting usage limits or shipping code with hidden defects, try the split:
- Plan in Opus. Get a detailed spec — architecture, file layout, acceptance criteria — before any code is written.
- Implement in Codex against that spec. Have it report at checkpoints/gates rather than running unsupervised end-to-end.
- Cross-review. After your primary model reviews its own code, point the other model at the same folder and ask for an independent code review. Treat what it finds as the real review.
- Keep everything in a folder on disk so both models share context with zero re-upload.
Related Insights
- Insight - Spec-Driven Parallel Work-Tree Development — Become the Spec Writer, Not the Coder — the spec is the handoff artifact between planner and executor; this adds the cross-vendor execution split.
- Insight - Multi-Model Debate as a Quality Control System for High-Stakes Work — cross-model code review is QC-by-disagreement applied to building.
- Insight - The Model Underneath Is the Multiplier, Not the Interface — choosing the right model per task (plan vs execute) is the multiplier in action.
Evolution Across Sessions
Builds on Insight - Spec-Driven Parallel Work-Tree Development — Become the Spec Writer, Not the Coder (2026-05-07) and Insight - Multi-Model Debate as a Quality Control System for High-Stakes Work (2026-02-26). New development: this session names a specific cross-vendor division of labor (Opus plans, Codex executes) plus the finding that a second model’s code review catches defects the first model’s self-review misses — with files-on-disk as the enabler for sharing context across vendors.