Getting Roasted by my LLM Style Editor

llm writing meta

I got caught being lazy. I wanted to experiment by writing a brief, fragmented post at the end of the day. The goal was simply to explain-brag that I’d found a neat trick to get an LLM to do a post-processing step that increased its output quality somewhat.

As is my routine, I sent the rough draft through my “Style Editor” subagent. That’s when it got personal.

Here are some selected lines from the LLM’s review of my prose:

  • “… reads like jargon explaining jargon”
  • “tepid business-speak”
  • “undermines its own point”
  • “the unfinished structure introduces sloppiness (ironically)”

But this was the kicker:

### Does This Need to Be Published?

Your "Sloppy Syntax" post already catalogs LLM output problems. This post adds the *action* of fixing them, which is valuable, but ask yourself: is this additive to your existing work, or redundant?

If you proceed, make it either comprehensively practical (show the full prompt, give more examples, make it a resource) or philosophically interesting (what does the need for deslopification tell us?). The current draft falls in the exhausting middle ground that you eloquently criticized in "Bernoulli Coding."

Damn. This is a far cry from the fulsome you’re-absolutely-right style feedback I’m used to. Honestly, it stung a little.

It’s also hilarious. I was not expecting to get roasted by the machine. I was unconsciously relying on a bit of the ol’ sycophancy to get me through to the next draft.

I get that this is me playing chess against myself: I’m using my own words to critique my own words. I attempted to approximate my aesthetic preferences, programming a strange machine to tell myself when I don’t like something that I’m doing.

There’s value in ruthless self-assessment, and in coming back to a piece in a different frame of mind from the one you were in during the initial creation. “Write drunk, edit sober.”

I’m used to this way of thinking while writing or making music, but this experience was new. Instead of coming back to the work with a different mindset, I’ve approximated the critical mindset. The shock came from having that receiving the editorial feedback while I was in a lazy mode of the creative mindset.

I’m thankful for it. The Style Editor shook me out of a complacent mindset and genuinely challenged me to put more effort into writing something worthwhile.

The trick, apparently, is stating your aesthetic preferences clearly enough that the system delivers tough love on its own. I told it what good writing looks like, fed it my existing posts, and it somehow decided I needed to be called out. The harshness emerged organically, which makes it more interesting than if I’d explicitly programmed it to roast me.

I’m still entertained by moments like this. It’s fun when these things surprise you, and it’s this feeling of playful
exploration that keeps me motivated to blend LLMs into my writing and coding practices.