Open-Sourced Codex Architecture Skill: Say No to Over-Engineering

Liparakakis · reddit · 2026-07-20

A developer noticed that AI coding assistants (like Codex) often fall into two patterns when giving architecture advice: blindly stacking unnecessary services and databases, or mindlessly suggesting a monolithic architecture. To fix this, they open-sourced a Codex skill with the core principle of "only adding complexity when constraints demand it."

The skill features three modes: GREENFIELD, EVOLUTION, and REVIEW. It refuses to pick a tech stack right away, instead interrogating the system's actual needs (required capabilities, invariants, failure consequences, hard/soft constraints) to build a minimum viable baseline, and performs adversarial checks to strip away redundant designs.

Additionally, the system tags every architectural statement (e.g., VERIFIED, ASSUMED, UNKNOWN) to prevent the model from quietly passing off guesses as facts. The author has also built an offline testing tool to compare blind test performances across different models and calls on the community to provide real-world test cases that can break the skill.

Original post →

More from coding & agent

coding & agent channel →