I made a tool to help manage all the claude chaos.
Rounds puts everything your Claude Code sessions make on one infinite canvas. Give Claude feedback directly, review changes, and see your history.
It all lives on your Mac. Local, private, free.
npx github:krispuckett/rounds
Have to have a Project / Thread level interaction and structure. Claude code does this exceptionally. The swarm ability will be supercharged with another layer abstraction above it that lets me operate as a human and see what’s going on, where I need to unblock things etc.
And please give us some refined and clean colors and icons.
Also - love swarm but the pixelated bees have zero correlation to “Anti-Gravity” and the other branding components.
Figure out the positioning of the brand and triple down on that.
A few words about Gemini 4 Neon.
At first, it seemed like a somewhat superficial model, but that impression turned out to be wrong. I suspect it’s because the model is relatively small, clearly smaller than Argon and Carbon.
But the more I interacted with it, the more pleasantly surprised I became by its unconventional depth of thought and intelligence. This model is genuinely smart.
I’ve also noticed that it’s less expressive than Gemini models tend to be, although that has usually been more typical of the Flash series.
In any case, Google should never have been written off.
Excited to launch Gemini 4 Neon: frontier-level intelligence, super fast task completion, and available to Pro & Ultra users now!! We think you’ll enjoy this model. Alchemy at its finest… 🧪
Odd UX Experience. I changed model when I ran out of quota and kicked off a swarm.
Came back to same chat next day. Changed model selector…did work…kicked off another swarm but it defaulted to using the model that was set previously for swarm 1 and did not use model that was in dropdown and currently in use.
AI psychosis is a weird time warp. You essentially teleport through the day and have this nagging idea that you aren’t doing enough all the time.
People that aren’t deep in it don’t even understand what it feels like to know the output you can generate at any time.
Sometimes you have to step away and reset. I do think there are only so many threads you can keep going at any given time and actually stay on top of them.
Part of the issue with AI is you as a human are the fundamental bottleneck and you just keep trying to power through it.
If you are 7x more productive with AI now it’s fine that you aren’t 9x more productive.
Slightly less than max is actually the better outcome.
The constant context switching. Sometimes you have to step away and reset. I do think there are only so many threads you can keep going at any given time and actually stay on top of them.
Part of the issue with AI is you as a human are the fundamental bottleneck and you just keep trying to power through it.
If you are 7x more productive with AI now it’s fine that you aren’t 9x more productive.
Slightly less than max is actually the better outcome.
@_lopopolo I feel like I have read every article I can find on software factories and “be more ambitious” / cut the models loose etc.
I still can’t fully understand the mechanics of how that actually works and what magic alchemy you cook up to get the end output to be durable and robust and continue to work through the ticket and test method Matt mentions.
Can you provide a deep dive or publish something (that you can externally) which actually explains and shows how this process works and can be done