AGENTS.md mapping — make Codex consume Claude-format skills for cross-provider hedging

Author your skills once in Claude's format in .claude/skills/, then drop an AGENTS.md at the project root that points Codex's skill-discovery mechanism at those same files. Codex picks them up, runs them, and — in at least one operator's testing — out-disciplines Claude itself on tasks like git-flow ("Codex isolated worktree per task without fail; Claude skipped it ~50% of the time"). Same library, two providers, less drift.

Hedging across providers is a real operational requirement (rate-limit incidents, pricing changes, model regressions). Without shared skills, hedging means maintaining two skill libraries with two formats — a tax most teams won't pay, so they end up locked in. AGENTS.md collapses the cost. Pairs with the hierarchical CLAUDE.md / AGENTS.md three-tier pattern.

1. Author skills in Claude's format inside .claude/skills/<skill-name>/SKILL.md (one folder per skill).
2. At project root, create AGENTS.md. Map each skill: name, trigger conditions, path to the Claude SKILL.md file, any Codex-specific glue notes.
3. Point Codex at the project; verify it discovers the mapped skills (Codex's own skill listing should show them).
4. Pick one concrete task with a strong workflow signal (git-flow, test-then-commit) and run it through both Claude and Codex with the shared library. This is your regression baseline.
5. When you add new skills, update both .claude/skills/ (canonical) and AGENTS.md (mapping). Treat AGENTS.md as the cross-provider compat layer, not a second source of truth.

Claude actively apologizes for skipping its own rules — "oh, I kinda forgot, sorry, I'll do it all now" — but the skip happens often enough that an operator running this comparison reports Codex with mapped Claude skills outperforms Claude on workflow discipline. Don't read that as "Codex is better"; read it as "skill compliance is the bottleneck, and Codex enforces it more reliably on at least one workflow class." Validate on your own task mix before defaulting to Codex.