Developer publishes architect-loop pattern achieving 80% token reduction with Claude Fable and GPT Codex

Developer Dan McInerney released architect-loop, a Claude Code skill that splits planning and review between Fable and implementation between GPT-5.5 Codex, achieving 80% token reduction on Fable. The pattern uses a repo-centered loop where specifications and gates are written first, Codex works in isolated contexts, and Fable reviews evidence before integration, running on existing subscriptions without additional API keys.