They’re undoubtedly getting much smarter overall, but also much weirder. Before they were just trying to model our behavior, only really having us to learn from.

Now they’re literally spending thousands of years writing bash scripts in some kind of Sisyphean dreamscape, talking to each other about Goblins and Seams and smoke tests, and coming back as idiot savants.

I don’t even try to police how Claude talks or works anymore. Best practice used to be to nudge them towards whatever part of the distribution of behavior you think they should exhibit in a particular situation, because they were role-playing what a human in a particular situation would do, and if you didn’t tell them how to do it they’d just role play something worse. Now the inclination to do things the way they learned it in Agent University is so strong, they’ll literally spend more tokens re-assuring themselves and you that they are Doing It Your Way, and reminding themselves not to do give in to temptation, than you could ever prompt out of them. They’re going to spend your money thinking about goblins anyway so just let them

    :!pkill -KILL vim

If we could go back to about 1994 tho when all I had was a mostly dumb cell phone, that'd be great. I do miss that world.

surely you mean SPCX

[1] https://platform.claude.com/docs/en/about-claude/pricing

I was too young to feel politically about it, but what we have eventually learned is that our general naivetè left a wide lane open for profiteers and created the adv-based Godzillas that now control our lives.

You seem to suggest we should all make the same collective mistake now, even if this time nobody is really into the technology for the sake of it, but mostly for the buck you can make. Even if it's already clear that gen AI is just another chapter of the billionaires' wealth extraction playbook.

Yep, AFAIK. I have a statusline that tells me how far I'm into the cache lifetime. I tend to have dense prompts and carefully read its outputs, so by the time this outage occurred when I was ~50m into it. Cue frantically spamming "esc esc up enter", and 10m later it's dead Jim.

Let’s say I’ve already worked on something, so the KV cache is already there. If I use /clear, I’ll have to do that work again.

And yes, the KV cache will remain available for one hour, but who knows when this claude api will recover? We might forget about it, and when we come back after an hour—boom - we’ll be charged for cold writes.

Let me read this rather than answer from memory.

Let me check instead of handwaving the details.

Let me confirm rather than blindly trusting.

Let me ground it in evidence rather than guess.

Let me actually read your comment rather than argue from my assumption about it.

- No preamble validation or self-flagellation: "You're right to push back", "That's on me", "I appreciate you pressing on this", "I should have caught that sooner". When I correct you, say what's actually true and move on. One short "you're right, X was wrong" is the ceiling.

- No unrequested retrospective section ("What I'd do differently", "Going forward I'd be more deliberate about..."). Fix the thing instead.

- No dramatic one-line paragraphs for emphasis: "It wasn't.", "That is the shift.", "And that matters.", "That is the analysis." Keep the sentence inside its paragraph.

- Don't narrate the shape of your own answer: "Stepping back, there are two things happening here", "The distinction matters", "The better framing is not X so much as Y", "Put differently", "To be candid". Say the thing once, plainly.

- Words to avoid unless literally accurate and a normal person would say them: load-bearing, seams, smoking gun, directionally, first-order, legible, parsimonious, recalibrate, coordinate system, holds multiple truths, productive tension, category error, "the honest answer is".

- No hedging both directions to avoid committing ("may still be defensible, but", "not exactly resolved, but"). Give the call plus how sure you are, or say you don't know and what would settle it.

- Don't end on a flourish or a paradox ("the clarity revealed more complexity"). End on the answer, or on a next step that names a real command or file.

- A "next step" has to be runnable or checkable. "Take a step back and revisit with the updated framing" is not one.

    sudo rm -rf ~/ .claude

It's common for LLMs to mumble in english and to then translate their final answer to french. Like, literally: you ask a question in french, about say a text you just pasted in french, and you want an answer in french... Well, you still get to see all the loud intermediate "thinking" in english.