A Rave Review of Superpowers (for Claude Code)
16 points by emschwartz
16 points by emschwartz
Why are there so many articles discovering that normal software development methods like architecture and requirements are useful for AIs?
It's definitely not a surprise that normal methods are useful for AI, but even if it's not a surprise, getting AI coding tools to follow those practices does make them a lot more useful.
I also rave about superpowers and tell everyone who will listen.
I just want to "teach" Claude to check its memories in every new session so that it won't repeat the same mistakes in every session.. Is there a "superpower" for this?
I've struggled with this as well (and am working on something in the area) - I think the current best advice is that you need to reference important memories in CLAUDE.md. If they are very important, transclude at least part of them there and link to them.
This problem also applies to skills. Claude doesn't trigger skills half aggressively enough. It's incredibly frustrating that you have to constantly remind it "you know how to search reddit, we have custom skills on that claude - stop googling with site:reddit.com"
There is no robust solve for this just yet. I’ve had success moving some things into hooks (I have a custom lint and similar). Claude doesn’t use skills enough but I’m sure they’re working on it.
I use both and I find that recent Claude is much more capable and that superpowers skills are now slowing me down.
Good to know! Do you start off in plan mode? I'll give stock Claude code another go
Depending if I start from scratch, or want to extend something, my strategy is to make it come with a plan, but it must put the plan in doc/plans (like superpowers), then we'll iterate over it. I try to make the plan on the high level surface (unlike superpowers) and break it in "sizeable" chunks that are commit in a branch.
Once this is done, then I enter in plan mode for each "chunk". Here are 2 projects where I used this technique: