Daniel Lowd

Imagine how fast it could be if you only rewrote it 10-20x more times!

Python -> Rust -> OCaml -> FORTRAN -> 8086 assembly -> PowerPC -> C++ -> Visual Basic -> C -> Perl -> Excel Macros -> Lisp -> Smalltalk

NumPy can inherit all the powers of all the languages!

OnePyToRuleThemAll!

And Claude will apologize and then we'll talk about how to avoid this happening in the future. ("Do you need another reminder in your CLAUDE . md file?)

Gemini and ChatGPT will sometimes insist that there is no chair while continuing to lean back in it.

That's more likely to happen with Gemini or ChatGPT, which can get stuck and double-down on obviously wrong things.

With Claude, it feels more like advising, or even parenting? Sometimes I have to say, "Hey, don't lean your chair back, remember? It's dangerous and could break the chair."

In my experience, working with Claude Code is a LOT like advising a grad student.

But my LLM experiences and my advising experiences may be different than yours.

And when Claude does something silly that seems like it should be preventable, I ask it how to prevent this. Would an instruction in CLAUDE . md have fixed this? Oh wait, there is an instruction there already? Why was it ignored? What will it take to be more consistent here?

etc.

YMMV

And then I have a discussion with first Claude before telling it to go forth and make the fixes.

Then back to second Claude and so on.

Some people automate this, but for research and research code there's less spec and more human guidance required I think.

"Hey, could you please take a look at HANDOFF_MAR6.md? I developed it with the help of a second Claude, and I think there's some good stuff in there. Review it carefully and let me know what makes sense and what doesn't make sense, because maybe second Claude made some mistakes too. Don't act yet."

...and then have a chat with second Claude about what's real and severe or not. And then put together a plan as a handoff document.

Then go back to first Claude and say something like,

"Hey, so, I have another Claude Code running but I'm pretty sure it's making some substantial mistakes. I'd like you to help audit everything it's doing. Please report on errors and oversights, big and small, but organize them by severity. Make a report in markdown. Once you're done we'll discuss."

I like to have a second Claude review and comment on the obvious stupid things done by the first Claude.

And I'm usually pretty verbose in my prompts, like...

Do not taunt Happy Fun Claude.

Yes. As someone directly impacted by these laws, I’m absolutely going to wring my hands over whether suicide policy actually prioritizes well-studied, life-saving prevention strategies rather than feel-good “solutions” that further marginalize and harm people already struggling w/suicidal ideation.

UPDATE: we now have THREE store bought cakes in our house.

(Also, cookie dough, but the cookie dough is homemade.)

Box plot showing prompt length distribution by week from November 2025 to March 2026, across all Claude sources (Code and Cowork, Desktop and Laptop), excluding messages over 500 words. Medians are consistently in the 40-75 word range most weeks. The heaviest weeks (Nov 24, Dec 01, Feb 09) show wide interquartile ranges extending up to 130-175 words with outliers reaching 400-500 words. Feb 23 has a notably low median of 21 words despite many outliers, reflecting short back-and-forth Cowork exchanges. Sample sizes range from 3 to 314 prompts per week, labeled above each box. Overall pattern: most prompts are under 100 words, but every active week has a long tail of 200-500 word messages.

My prompts tend to be long, too. I have this habit of giving Claude several paragraphs of rambling for full context.

10% of my Claude code/cowork prompts are over 200 words long. (And that's excluding all 500+ word prompts, which often contain cut-and-pasted text.)

Bar chart showing words typed to Claude (Code and Cowork) per week from November 2025 to March 2026, across two machines (Desktop and Laptop). The top chart shows word volume: early usage was Desktop-heavy at 20-25K words/week in late November, tapering off through December-January. A large spike of ~33K words appears in early February, dominated by Laptop Cowork usage. Recent weeks show 8-13K words with a mix of all four sources. Total typed: 125,217 words over 100 days, excluding 103,524 words of pasted bulk content. The bottom chart shows prompt counts with similar patterns: ~300 prompts/week at peak, with the February Cowork spike being especially prompt-heavy. Total: 1,585 prompts. Colors: red (Desktop Code), orange (Desktop Cowork), blue (Laptop Code), teal (Laptop Cowork).

I've typed over 125,000 words into Claude Code and Cowork over the last 100 days.

(Figure and alt text produced by Claude Code. Naturally.)

I knew working on suicide research was going to be heavy.

But what's actually starting to get to me is reading all of these chatbot suicide laws and legislative proposals that call for measures that have been shown to exacerbate crisis.

www.governor.ny.gov/sites/defaul...

"Computer, create an adversary capable of defeating Data" is the first GenAI jailbreak.

"Computer, create an adversary capable of defeating Data" is the first GenAI jailbreak.

Replicators have a list of “recipes” and each one is already programmed.

In the holodeck however, you can prompt the computer to build you any 3D interactive world that you want, including all the rich characters inside it.

It’s very much going from a prompt of a few words to a usable experience.

Yup, samesies.

Also, it thinks this is one of my most negative valence posts:
"It’s so weird that Valentine’s Day is never on Friday the 13th."

And it thinks this is positive, when it's arguably a rather dark sentiment in context:
"It makes me extra glad I have tenure."

So my experiences haven't been all bad. But the meds were way more important than the talking.

And when he asked, "Do you need to compare yourself to everyone else?"

And I said, "Yeah, it's kinda my job, that's how they decide whether or not to fire me."

...he just let it be and didn't try to argue me out of it. Didn't try to fix the things that couldn't be fixed. Which surprised me.

I did have one guy who was somewhat helpful, because I was worried (with good reason!) about not getting tenure, and he had a PhD and lots of friends in academia. So he had relevant perspective on my particular situation.

When he said, "You'll get tenure," it had some weight to it.

My experience was more that I would be open and vulnerable but receive very little guidance or support, so I'd just leave each session feeling worse than before. (Not always, but often enough.)

I guess I need my therapist to be a bit more sycophantic to be effective?

Seeing a human therapist requires waiting 6 months and maybe spending $1000 for a few visits (depending on your insurance and if your need qualifies for it).

...and then doing it again when you realize that the first therapist wasn't a good fit.

When I log into my Xenix system with my 110 baud teletype, both codex *and* gemini are just too damn slow...

This whole interchange is peak AI.

AI is amazing, and it does amazing things, and it understands things remarkably well, and it also does things wrong, and if you're not paying attention you may miss that it's wrong.

“It’s a consciousness problem, Michael. How hard could it be?”

If your job was to know when to hit the semicolon key on the keyboard, I'm sad to say that's not an employable skill anymore.

If you shape ideas into tools and art via code, it's really so much better than ever now, and I don't see that changing soon.

Oh interesting, people who don’t know how to build software are getting mad at my post about building software. Cute.

Let me be clear, over the next year, the job of software engineer will shift dramatically to no longer have typing syntax into an editor as its primary time sink.