I'm wondering if Bun would be a good embedded runtime for Claude to think in. If it does sandboxing, or if they can add sandboxing, then they can standardize on a language and runtime for Claude Code and Claude Desktop and bake it into training like they do with other agentic things like tool calls. It'd be too risky to do unless they owned the runtime.
Another problem I inadvertently dodged by using Jujutsu with Claude Code :)
I tend to send a lone "commit" message to Claude when I think I'm in a spot I may want to return to in the future, in case the current path doesn't work out. Then Claude commits it with a decent message. It knows how to use jj well enough for most things. Then it's really easy to jj new back to a previous change and try again.
There are still branches, but they aren't named by default. You give them names with "bookmarks", which you can push to remote git repositories as branches.
This lets you work on things without having to worry about giving it a name. This turns out to be pretty helpful when you're experimenting — just "jj new <revision>" and start editing. If it turns out to be something you want to share, "jj bookmark create <name>" and then you can push it. (You can also push without giving it a name, in which case you'll get a git branch with a name based off of the change id.)
Change IDs stay constant with each change, so you use those as a type of branch name when switching between the features you're working on.
A change ID is stable over time as you tweak the message of the change or the files edited by the change. Each of these changes become a new immutable git commit under the hood.
The fact that change ID is stable is very convenient for humans - means you have something explicit to hold on to as everything else may change over time.
The way maud lets you compose markup works very nicely with htmx. The HX-Request header lets you know if the request is coming from htmx or if it is a regular request. You either call the top-level function if it's a regular request to get the entire page rendered, or call a subset of functions to get the appropriate partial rendered if it's an htmx request.
It's also nice to easily have tests for the rendered pages. My unit tests cover verification of the rendered HTML, too.
GitHub will call the webhook after a push to main and a successful test suite run. Snare runs a shell script on my server to git pull, build, deploy, and call a cronitor.io hook for monitoring deploy success.
I've been pretty happy with how relatively simple it is and how well it works.
> I have a pet theory about the "AI revolution" or AGI that keeps getting relentlessly confirmed as events unfold: Microsoft sees a massive financial upside to this technology that no one else sees and this is being kept under wraps.
If AGI is "a highly autonomous system that outperforms humans at most economically valuable work", I am Microsoft and have AGI, and other businesses do not, I am putting it charge of a Windows VM in Azure and offering it to companies to run aspects of their businesses. Why stop at "GPTs" if I can offer you specialized Clippys?
Put all your data in Microsoft 365, let Clippy do its thing, and you're saving a lot of money on not supporting people. Microsoft gets their cut, and you get to fire your employees. Win-win.
Kagi allows you to provide custom CSS if you want to tweak it to your liking. I've used it to remove some of the widgets I don't personally find useful, and to replace some icons.
I bought a Flowbee in 2009 for $83 total. I cut my hair around once every 3 weeks, so I have paid 34 cents per cut so far, ignoring electricity cost of running a vacuum and the Flowbee for a few minutes.
