Markdown instructions to make Claude use Codex subagents

Ansh Nanda shares a Markdown/CLAUDE.md pattern for making Claude act as the planning and discussion layer while routing coding, discovery, implementation, research, and other token-heavy work to Codex or subagents through a skill. Why it matters: The signal is not the exact command or install link; it is the workflow contract. A short Markdown instruction file can turn Claude into a planner/router and force execution into Codex-backed subagents, reducing expensive context burn while keeping Claude focused on strategy and review.

Try this: Adapt the idea safely: write a reviewed CLAUDE.md rule that says Claude plans, critiques, and reviews, while Codex/subagents handle implementation and research. Measure token use, latency, test pass rate, and review quality before trusting the claimed savings. Do not run the linked commands or install the linked skill without inspecting them first.