With the amount of programming that’s getting done “while I slept,” you’d think AI is making us all really well rested…
10.02.2026 12:48
👍 1
🔁 0
💬 0
📌 0
With the amount of programming that’s getting done “while I slept,” you’d think AI is making us all really well rested…
Just ran across this post by @andrewnez.bsky.social that puts words to my deep feeling of betrayal by Ruby Central in their awful governing of rubygems. "Package Registries Are Governance Providers," meaning their failure betrays the core promise of a package registry. nesbitt.io/2025/12/22/p...
These aren’t just my gray dish towels. These are the gray dish towels in the world.
Skillet can now start up with an existing claude task list.
$ skillet --task-list=mytasklist --prompt "Complete the assigned tasks"
You can have another claude create and assign these tasks to a skillet worker (I think we could improve the interface though)
github.com/martinemde/s...
What sad is that a mass migration to a competitor is so unrealistic, and yet I still agree.
And another thing! GitHub missed probably the one major thing we really needed them to do: separate AI actions from human actions. The "fine-grained" tokens are woefully inadequate for a situation where agents act on our behalf. GitHub ignored critical security controls to give us mediocre Copilot.
It's so terrible. Things that used to be sync is now eventually consistent, and I'm using "consistent" loosely. More crashes and broken pages than ever.
Claude Code merged the duplicative slash commands into skills, and made skills better in the process. Happy to see this right as I was working on skills a bunch (the split was confusing, and a merger seemed inevitable). martinemde.com/blog/claude-...
Why not both? It’s scriptable claude which was always possible but so ergonomically bad that you have to vibe something up every time and it’s always custom and never very good. The point is to easily elevate claude into the shell so it does a canned task in CI or in a simple bash for loop.
You’ll never have to deal with it again. Problem solved. New problems created but different!
Lately it feels like the fan is just blasting at turbo speed and the ducts are full of shit but the fan is insane and the shit is hitting it all the time and the fan does not give a shit.
Delete the app. 200 is too many. Kill it with fire.
You can delete messages, right? There’s no hope.
I’ve written countless throw-away headless claude scripts.
- None work on the first try
- Ugly prompts get buried in the ugly script
- Permissions wrong, CLI flags wrong, nothing works
- Output choices: Nothing / JSON flood.
Agent Skills + Skillet fixes all of this: github.com/martinemde/s...
i think a core beef that i have with a lot of this "automation is bad" sentiment lately is that there is no inherent moral good in laboring
the story of humanity is one of invention, where we improve our conditions by building things that help us do more things more easily
New focus & blur shaders in Ghostty enabled by my recent PR. I made a little "emulator" for them on my site where you can test them out.
I've been absolutely nerd sniped by these lately. martinemde.com/blog/ghostty...
Hmm, I would not describe it that way. It’s about running Claude as cattle, not pets, and using hierarchical project management and a special merge queue agent to integrate everything, overseen by a high level orchestrator. All continuous, no halting, not lost state.
Yep. There’s no way the future is “my mayor bungled the polecat sling and the deacon kicked boot and screwed up my rig.” But the agent managing workers with organized small work streams that get refined and checked, probably.
At least it’s fun.
With Steve Yegge’s Gastown, we have a new mantra.
Move fast with broken things
Love it.
github.com/steveyegge/g...
Will take a look. Lately I’ve been kicking gastown’s tires. It’s mostly broken but with cute animals that do work in a magical black box way that works well. Still getting the hang of it.
Oh totally. Sounds like you’re doing with token numbers what I’m using the percentage for. Either way you kind of need to get the feel for the agents reliability as the context grows. That’s why I added the percentage and colored progress bar. Thanks for the link!
I can see number of tokens. Do you have a percentage I’m missing or are you just familiar with the number of tokens that means your closer to context rot territory or compaction?
I should try it out. Thanks for mentioning it. I’m smitten with ghostty right now though. Custom shaders are super fun.
Do you happen to know if the progress bar codes work in Alacrity as well? I’m curious what it looks like.
Maybe tmux is partly to blame? I'm not certain. I was never a heavy tmux user by I did set up ghostty to map to tmux bindings (with a different prefix) so that I don't use tmux for the most part. github.com/martinemde/d...
Oh interesting. I’ve seen ghostty go wild on RAM but my understanding is that it’s a bug. I’ve also been skeptical that’s it’s not just Tahoe, which is when this all started. (I’ve seen ghostty “use” 600GB of RAM which is obviously not actually happening when my combined SSD+RAM is <536GB)
Here's a screenshot. It's fully customizable because it _is_ starship. Notice my context is about 20% full. Very useful at a glance.
I made a claude code statusline that updates Ghostty's progress bar (that blue line at the top) as you fill up context.
I know there's a lot of these out there, but this one uses starship directly for rendering, which was somewhat rare when I looked.
github.com/martinemde/s...
Yay congrats!
Interesting approach: Suppress Josef's comment, then suppress anyone else that isn't in support of the new "governance."
Making a deal with the person that broke the last deal is futile, but what do you achieve by suppressing the people saying that?
Suppression just brings more attention.
