From the first prompt, I knew GPT-5 was different. Not trying to hype it up, just calling it like I see it — it actually follows my instructions the way I want, which hasn’t always been the case with other models.
I’ve been building almost exclusively with RedwoodSDK lately (their new SDK on top of Vite, RSC, and Cloudflare). Most models I’ve used default back to their training defaults — for full-stack that usually means Next.js and REST APIs — but GPT-5 doesn’t do that. It’s like a breath of fresh air.
Now for the real reason I’m posting:
I want to customise Codex to fit how I actually work. My workflow has a certain rhythm to it, and I’ve noticed Codex doesn’t have the same sub-agent and workflow-type features you find in something like Claude Code.
Anyone here figured out how to:
-
Run sub-agents inside Codex or something similar
-
Build proper workflows that stick to your dev style
-
Or just hack Codex to feel more “yours”
If you’ve got tips, resources, or weird experiments that worked for you — I’m all ears.
Big thanks to the OpenAI team — GPT-5 is honestly making building fun again