The New York City subway map is so iconic you can build an identity around it. You’ve seen it, or its 1979 redesign by Michael Hertz, printed on coffee mugs and t-shirts and chico bags.

After twenty years of dedicated tmux use, I listened to my younger colleague and tried zellij: a screen multiplexer written in the hot new language Rust.

After 3 months of use, I am going back to tmux.

Pros:

• On-screen menus make key bindings far more discoverable than tmux’s man page ever did
• Mouse support (but this is actually a con for me)

Cons:

• Its key-combo design clashes with the command-line tools I use daily (although it is fixable by drafting a custom config)
• Mouse support
• Sine qua non It doesn’t have a stack of named copy buffers

Back in April, word got out about an unreleased, highly capable version of Claude called Mythos that, when pointed at large codebases, could surface serious, previously undiscovered vulnerabilities — including in software that was decades old. It was good news copy and the press fanned flames for several news cycles.

But after having done a lot of AI-assisted engineering through Summer 2025 and doing a lot of AI-assisted process management at work in 2026, I have to ask:

If the current models and tools are Ragnarok-grade deadly, why isn’t it more visible doing basic engineering?

Or:

If the current models and tools are so sophisticated e.g. Claude Code for Design, why is Claude Code on the desktop so damned ugly both aesthetically and engineering-wise?

In short, there’s a lot of visible lack of dogfooding of the AI tools that makes their claims look hollow. I’m not saying the emperor has no clothes, but at this point he’s doing a hot monarch summer in lamé booty-shorts and a crop top. The bungling is so obvious it makes me wonder what level of dysfunction has beset Anthropic’s product management and marketing.¹

Here’s a list of bungled market communication.