I've found that these models are extremely good at reasoning since late last year, but they're basically 200 IQ Amelia Bedelias. So my working model earlier this year was to give it a brief instruction and watch how it failed, which was generally a lack of context that it couldn't have possibly known. Then I'd add that context to the team AI docs repo, try again from scratch, and iterate until it did what I wanted without me having to give a bunch of detail in the prompt. Over the course of 2-3 months, I needed to give new context and steering less and less.

I also encouraged my team to paste any inquiries or alerts we got in our slack channel verbatim into codex and follow that procedure until it could figure it out as a one-shot.

I still treat it as more of a design and coding partner day-to-day, so I'm typically not trying to one-shot, but it can e.g. be triggered by a pagerduty alert, reference our code, check grafana panels, query application state across different servers/clusters, and come up with bug fix PRs all autonomously, and then post its analysis, mitigation suggestions, and PR link to slack for the on-caller to review. This was all just some docs, scripts, and a little listener service to trigger a one-shot model prompt that we built.

As a design partner, it's made it easier to add more telemetry or create rapid prototypes to check my assumptions. I've encouraged my team to regularly ponder on what facts they wish they knew, no matter how difficult it would be to find out. Pretend you could just ask God anything you want to know. Then ask the AI to go add whatever's needed to find out. We should all move toward a perfect intuition for what our system is doing because all questions can now be easily answered if we can only think to ask them.