Secondhand embarrassment

Speaking of psychology … a weird expe­ri­ence last night, as I found myself with a couple of spare hours and decided to finally try a Claude-assisted writing exper­i­ment I’d been imag­ining for some time.

I had the whole project organized: a bunch of sup­porting documents, an elab­o­rate prompt for Claude Code (this was part of the exper­i­ment, to use Claude Code’s looping amend­ments rather than the big GENERATE THIS of the API), a little Ruby script to kick it all off.

Yet I found myself hes­i­tating to run the script. It was just the lightest reluctance, enough to send me back, again and again, to fid­dling with the prompt, adding some notes to the sup­porting documents.

I realized: I knew Claude would do a bad job, and I was feeling a kind of pre­emp­tive sec­ond­hand embarrassment. Or I wanted to avoid that feeling. Or some­thing.

Sec­ond­hand embarrassment: regarding the cheery, con­fi­dent output of a com­puter program. Wild stuff.

I got over it, obviously, and ran the script, and Claude Code charged ahead on its own for quite some time — I should have recorded how long — and when I returned to the terminal, there was a sub­stan­tial piece of writing waiting for me, and it was: bad.

(Standard caveat: from the per­spec­tive of 2016, when I started tin­kering with this stuff, the writing was amazing. But/and we are talking about real work here, not sparkly demos. The ques­tion is not “has this tech­nology advanced in breath­taking ways?”—it’s “can this tech­nology (yet) sup­port world-class cre­ative work?”)

My exper­i­ment was built around a light­weight method for pro­viding notes and cor­rec­tions (in my imagination, some­thing I could do in spare moments, on the go) then feeding them back through, an elab­o­ra­tion of my trial back in May 2025. But I saw clearly that it wouldn’t work. The required cor­rec­tions apply to every scale, from macro struc­ture to word choice, and they are so exten­sive they amount basi­cally to: writing.

So … just write it.

Maybe that’s the difference, in the end, between prose and code: even world-class code con­tains boilerplate, familiar incantations, throat-clearing. World-class prose absolutely does not — and indeed a huge part of its value is that there has never been any other writing like it.

For the curious, some details:

• I used Claude Code in its head­less mode

• I selected the latest Opus model

• I fed in QUITE A LOT of context, whole books of ref­er­ence mate­rial — the most I’ve ever dared

• Nei­ther my direct guid­ance nor the ref­er­ence mate­rial seemed to do any good in terms of tone and style. In my assessment, the Claude models are get­ting worse at writing with any verve as they get better at struc­tured programming. This makes sense; the “style” of the rea­soning traces is chirpy and functional. I acknowl­edge there might be spe­cial guid­ance lurking in the Claude Code-specific setup that’s working against me, too.

• I will observe, echoing many Claude Code users, that the tool’s appetite for clear instruc­tion pro­vides a healthy prompt: what am I doing here? What do I want? So, my time wasn’t totally wasted last night; I’m happy to take some of the guid­ance I pre­pared for Claude and use it myself.

There are going to be a lot of reluc­tant script exe­cu­tions and heavy button presses in this new era. You think about what’s going to happen, and you say, yeah … go ahead.