2026.10.29

Merge Codex and Opus implementation plans, and use Codex for OpenAI API docs

Have both models draft plans at max reasoning, compare them blind, then fold Opus's strengths into a Codex base.

When building a complex feature, I recommend having both Codex and Opus produce implementation plans in Markdown with reasoning at full max, then merging the two plans.

When comparing them, place the two files with the authors hidden and have them evaluated. Codex’s plan is usually chosen…(evaluations often say it fits the existing design, or that it will work more accurately)

Then I take Codex’s plan as the base and fold in the good parts of Opus’s plan, plus the finer technical and architecture details. What you get is something with the unnecessary over-engineering cut out.

When I actually compare the two, there are plenty of moments where I think, “Codex’s plan alone is not quite enough, Opus’s point is more correct,” so combining the two implementation plans seems best for now.

Also personally, I recommend using Codex when you use OpenAI’s API. Through skills and plugins it can read OpenAI’s official specs and docs directly, so I have the impression there is less hallucination.

Claude Code cannot read those. For getting the right information and gathering context, it seems better to use the models and tools that are good at each…

https://x.com/suna_gaku/status/2115648808222953704