decided to open source my carefully tuned Codex set of skills, subagents, AGENTS.md. The core is the heavily opinionated workflow that agent is forced to follow. Burns more tokens but the output is much better. Polished by heavy use in multiple commercial projects. There's also a set of tools to keep it in sync across multiple machines + a code-server-based control surface app