weka

I've been treating AI work from the beginning as the work of an intern, and that has worked out pretty well. Sometimes if the feature is complex I first do a pass of light scaffolding where i start to structure the code how I want it, and leave placeholders. LLMs thrives around structure, as they are designed to recognize patterns. From my experience, if you design a whole new feature and the AI can't really find proper structure around it in your codebase, result quality will be much lower. I then make Codex do most of the implementation work, and do a very light review myself, to make sure the code is not unnecessarily bloated. If codex doesn't structure the code the way I imagined and I can think of a more efficient way to do it, I'll guide it. Then in a new session I ask codex to review the changes compared to the base branch, most of the time he will find issues, in which case I make him fix what's worth fixing, always making sure the fix doesn't lead to unnecessary bloat through a very light review I then trigger a new review in a new session, and keep repeating this until there are no more issue worth fixing. When codex is done fixing issues, I then usually make Claude do a pass of review. Claude is usually much more nitpicky and often a lot of his findings are not really worth fixing. So I feed the output to Codex, and ask it to review what is valid and worth fixing. In some case, it will highlight a valid problem than Claude found and Codex missed. Once all those steps are completed and both Codex and Claude don't find any meaningful issue anymore, I'll do a manual very in-depth review myself, in the same manner I would review the PR of an intern. Then another person from the team will also do another pass of manual review on my PR before it is merged. Ultimately, I think it also depends on what type of software you are working. I could probably live with few less pass of reviews that I am currently doing if I wasnt working on a financial app like @KyanExchange where there is no room for mistake.