Why Codex works better than Claude Code for my production monolith
The author shares their experience using Codex and Claude Code on a long-standing, complex Python monolith, finding Codex better suited for backend tasks due to its adherence to harness-engineering principles and superior context handling. While Claude Code performs well on UI-heavy tasks, it often fails to reuse existing tools and requires more iterative corrections for architectural decisions. The preference for Codex stems from its planning ability and reduced need for explicit guidance in a layered, legacy-rich codebase.
Opening excerpt (first ~120 words) tap to expand
Over the last year I mostly used Codex, but during the last month I tried Claude Code with Opus 4.6 and 4.7. These are my notes.This is not a benchmark. It is just my experience from daily use on one production codebase. For some medium-complexity tasks, I also ran both tools with the same prompts, but I did not try to make this a controlled evaluation.TL;DR: for my production Python monolith, I still prefer Codex.The codebase is a many-years-old Python backend. It has several architectural layers from different periods: a newer experimental DDD-ish style, older but still well-structured legacy code, and very old fragile spaghetti code.We usually do not rewrite old parts unless we have to. The preferred strategy is to leave them alone until they are naturally replaced or removed.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Ycombinator.