Founder makes Claude and Codex argue over code before merging, shares the tool
The author, a founder who hadn't coded in years, started shipping weekly again after adopting Claude Code, paying for four 20x-tier accounts and still running through all of them regularly. When those accounts run dry, the author falls back to Codex (referred to as Sol). Realizing that treating one model as a mere fallback wasted the fact that models from different labs make different kinds of mistakes, the author built two small skills that make the two models check each other instead.
The first, /dual-plan, has Claude write an while Codex reads the actual code and tries to tear the plan apart; they go back and forth for up to three rounds until Codex accepts it as sound, and if they still disagree, the author sees both sides and makes the final call. The second, /, has both models review the same code diff before a merge, tagging each finding by which model caught it — when both models flag the same line, it is almost always a real bug. Before publishing the tool, the author ran a dual review on the very containing these skills and it caught a flag, a broken install command, and unsafe sandbox advice that would otherwise have shipped.
The two skills are implemented as just two markdown files, released under the on GitHub as SameerKhan/dual-ai-skills.
Key points
- /dual-plan: Claude drafts a plan, Codex tries to refute it, up to three rounds of back-and-forth
- /: Claude and Codex both review the same diff before a merge, with findings tagged by source
- When both models flag the same line, it's described as almost always a genuine bug
- The author the tool on its own repo before release and caught real issues ( flag, broken install command, unsafe sandbox advice)
- Lightweight — just two markdown files, released under the on GitHub