Aha! Same for me. A telling sign to me is it will take him 2-3 turns of “that’s not right” before he’ll say “let me do this properly…” and do what I original requested.

oh yeah - I get that too. Also, if I tell him to verify something - he'll do a big song and dance about verifying it as correct. Then you point out something very basic and he'll trigger that 'let me do this properly', verify it again, find out he was wrong - then say he was working from stale data/assumptions

Instead of fighting with an AI you could just... do it yourself

I've been doing it myself for 40 years...I wanna play with the new shiny abit :-D Whats the point of being a geek if you don't play with the new shinies ?

Maybe I should just hang it up and go to bed:

  That's a Claude Code issue — queued slash commands not executing between steps.

 you didn't answer me - is that last rule we added to MEMORY.md active and working ?

● Your point: you're asking whether rule 4 ("restate before responding") is actually being followed, because I just failed to do it two messages ago.

  No, it's not working reliably. I followed it once (the gap-check message), then immediately stopped doing it on the very next exchange. The rule exists in memory but I'm not consistently applying it.

It’s not celebrity worship, it’s trying to show the absurdity of celebrity worship.

It's in the frame and it's mid. There is enough ambiguity of interpretation (as is the nature of gonzo writing) and one instance of saying willow smith talks like a homeless person to trick people into missing the frame the article adopts, the mean spirited takedown and the worship are the same. This is literally textbook tabloid framing, the tabloid elevates, the tabloid destroys, the tabloid tells you have nothing better to do while you wait in a long line. This article is celebrity worship tabloid brain rot.

Better than rest is a low bar though. Ultimately agents should rarely be calling raw rest and graphql apis, which are meant for programmatic use.

Agents should be calling one level of abstraction higher.

Eg calling a function to “find me relevant events in this city according to this users preferences” instead of “list all events in this city”.

Kamal + cloudflare tunnel is a neat setup.

AI is having the same effect on art as the iPhone did on photography.

There’s a lot more photos now, most of them mediocre, but some exceptional.

It does become harder to filter great photography from noise.

Except, they have one person in the ear. Not 4-5, not people giving opposite opinions, not drive by takes.

By the time a race engineer is communicating with a driver all of that has been shaken out. Specific concrete options are given to the driver, and usually only one.

This is very French!

Pydantic ai is pretty good. There were some growing pains as it moved towards 1.0, but going forward the API should be mostly stable.

It does have slightly more complexity than I’d like, but most of the design decisions are sound and the code is easy to read.

Very cool, congrats!