It is a strange time. I personally have seen a lot of people produce very high quality code using AI - code that is well factored and tested. I have repeatedly experienced AI producing higher code quality than what many developers can produce. At the same time I see many people post about having the opposite experience. It is very hard to understand why people can have such polar opposite experiences.
I worked with three such posters who had a bad experience that were close friends of mine. In all three cases we discovered that the tools they were using were suboptimal and that they were not configured to provide anything even close to the best outcome.
Because of this experience, I decided to write this blog post. In the post I walk you through the tools I use, and more importantly, how I configure them. In my view, this makes a huge difference in the quality of code that is produced.
This is a quote from the post → “Think of the agent as a fast car without a steering wheel and brakes. Your job is to tailor it for your specific workflow and needs.”
Really nice write-up, I enjoyed it. I plan to take inspiration from some of your workflow items, thank you.
In my experience the results from LLMs are actually fairly deterministic at this point: the more details and guard rails and constraints you give it, the better it does. Get lazy and start using vague language and have it just stumble upon wall to wall and say it “fix it!”, the worse it gets.
Agree with @dimitarvp - really useful writeup. I’ve been using Cursor interactively a lot over the last year or so, but been struggling to keep Claude Code & Codex on the straight and narrow. I had set aside some time to go through Mitchell Hashimoto’s recipe here: My AI Adoption Journey – Mitchell Hashimoto, but your notes will really help accelerate the process - your timing couldn’t have been better! Thanks.
I’ve been using your config for about a week now and I can say that it did indeed improve the quality of the code claude gives me.
But, at the same time, It seems like it also is using way more tokens than before, I basically can’t use it for more than a few minutes and I hit my 5 hour rate limit… (I use the pro plan btw)
Do you have the same issue? Any suggestion on how to improve that? I tried installing context mode plugin since it should help with token usage, but so far it didn’t help much (if at all).
Two things you can try to reduce the tokens it is using. You can set the CLAUDE_CODE_EFFORT_LEVEL value to “medium”. Also, I just updated my usage_rules which required all new configuration - but it moved a lot of things from the AGENTS.md file to a new Skill. This will reduce the context. Give this a try and see if it helps.
Well I have not used it and I don’t want to slander it (I am actually very curious to try it but have not been working close to the frontend part of Phoenix projects in a while now), I am only saying what I remembered people saying in its announcement thread.
It’s very possible this has been fixed and made better in the meantime.
