What Is Wrong With Ur Claude
If you're mildly dissatisfied with software, WHAT IS WRONG WITH UR CLAUDE
it has gotten super common for online peeps to muse, "it would be nicer if your product did [X] and [Y], instead of [W] and [Z]"
esp. in web3 where being early means users put up with warts in return for providing fast feedback loops
and growth and go-to-market gives power to a group of users between the highly-technical and the "normie" non-technical group. traditionally, this was professional project management. now, it's idea-guy founders with exquisite taste, and an engaged and vocal community that loves your product
but the value here is a willingness to see things like a stupid person. and to speak up with a certain sense of entitlement.
more charitably, you want advocates of The Principle of Least Surprise
AGENTS DO THIS NATIVELY and can be tuned down in how annoying they are, with a touch of sycophancy and a sprinkling of anti-clanker prejudice
I'm having so much fun and agency right now, with well-structured open-source projects that are very much AI-enabled, but extremely low in bloat. I can add my dashes of slop in there to try out stuff, without fearing the whole thing will collapse under the weight of slop.
Two examples:
example 1: Kepler
I took my friend Steve's featherweight Go binary, added a system-setting-aware Light mode for my aging eyes, and tweaked the list into a sortable table.
If I was hand building this whole thing, these tweaks would be trivial, if and only if they were layered on top of prior technical decisions that overlap with my comfort zone: React/Svelte, TanStack/Shadcn/Daisy, I don't want to ever touch Caddy, etc.
In this case, Steve used a stack I'm only vaguely aware of. I could tell it was built solidly, without knowing exactly how. But I would have had to chew glass before feeling confident in making trivial changes.
Temu Claude figured out what I wanted, presented choices, and walked me through all hiccups in deploying on my specific set of hardware and software.
example 2: Quilcast
This is a much more complex project. Nikita basically had his agents reverse engineer and speed-build Warpcast, which at one point was a product that raised $180m, valued at over a billion dollars. Plus a bunch of unhinged experiments.
But the code benefits from Borodutch's many iterations, and the agents have written it with 100% no-bloat.
I can trivially self-host it, and add my own little experiments. Like detailed spam labels.
