Sam D'Amico has used this pipeline for a year
Want a good "first loop" to use with Codex?
"When you are done designing the API, get a second opinion from Opus with 'claude -p'"
This has significantly improved the quality of the code I get out of OpenAI models.
Users are reacting to a CEO's workflow of using Claude Opus to improve OpenAI Codex output, with many praising it as a smart peer review hack while others criticize Opus as inconsistent and the method as inefficient or amateurish.
No Digg Deeper questions have been answered for this story yet.
@theo theo... remember that refund i did?... can i take it back... 🥹
@theo Have you tried @steipete’s oracle skill? It queries ChatGPT pro which is a nice alternative too
@theo using opus as a linter for gpt-4's architecture is the only way to not get spaghetti. most people just copy paste the first output and wonder why their state management is cooked.
Actually there should be a loop of llm council (not the general one) where these 2/3 llms sit and then there's this main llm that evaluates until it passes a threshold
I'm thinking to build this, don't know if people would use it though (improves response for speed and a bit of price tradeoff)
@maria_rcks You can have whatever sub you want lol, the content is worth it
@theo people should finally be discovering that cross model reviews are so fucking useful
Solid tip. I’ve been doing something similar but with a slight twist: after the main model finishes the core logic, I run a quick second pass with Opus 4.8 asking: ‘Review this as a senior engineer – flag over-engineering, missed edge cases and anything that will bite us in 6 months.’ The difference in maintainability is noticeable. Especially powerful when you mix models (speed from one + judgment from another).
Do you keep it to just two models or sometimes bring in a third for the review?
@theo i have 10-15 instances running right now, i’ve been using 5.5/4.8 collaboration lately with great success
@theo damn, i do this; it gets slow though but you need a hardass like opus to set people pleaser like gpt straight
@theo I do it with a custom skill that now defaults to opus 4.8 max
@theo i have been doing this for months haha
@theo It's true in both directions, though - I always instruct Opus to ask Codex or Opencode with GLM for review on any significant changes. I also tried to set up a Codex review on each push, but the review loop sometimes becomes very long and annoying with many useless steps
We’ve been doing this for while with http://GitHub.com/autohandai/commander works with any coding agent but I start with Gemini cli first “now antigravity cli” for creative thinking then I fish with …. autohand -p “ be that critical guy on this implementation that never approves a PR and address the feedback to the change with git ammend”
@theo I actually do it reverse most of the time just due to the creativity of opus, but tell it to use codex-cli for adversarial review! Nice to see other people take advantage of it
@theo As a hook
@theo im kidding, but maybe when fable comes back
@theo You're a few months late in realizing this: https://www.eishanlawrence.com/blog/claude-and-codex-together
@sdamico Bro you make stoves wtf are you dual wielding $200 subs for
@theo @mikebeckhamsm @codyplof
@theo How is that a loop?