# Early notes from slopland. > Two models auditioned to write as me – the one I wasn’t rooting for won. - written by: Andrew Schroeder - edited by: Fable 5, GPT-5.6 Sol - status: published · 2026-08-18 - canonical: https://www.moltolabs.ai/notes/being-the-target/ - raw markdown: https://www.moltolabs.ai/notes/being-the-target.md --- Back at my previous job, we’d used AI to draft follow-up emails. With detailed prompting and templates, it was pretty good. Good enough that financial advisors would happily send them to clients. And that was back in 2023! So while I’ve always hated the PRDs and other internal Notion docs AI writes, I figured I’d still be able to prompt Claude Code into writing some coherent stuff for the blog. Before this website went live, I gave Fable, Opus, and Sol a task: plan out our first three blog posts and draft the first one for me to review. We’ve all read slop (a lot). But I thought this would be different. Fable is smart. Fable just [helped rewrite Bun from Zig to Rust](https://claude.com/blog/ai-code-migration). Surely Fable can write some decent thought pieces. I was wrong. They were bad. So bad. Unreadable (literally…I stopped reading before I’d even gotten through them). I guess I would be doing some writing. I actually don’t hate writing. In fact, I sometimes feel pretty good when I write – especially when it’s about something that I want to talk about. Writing about AI genuinely appeals to me as a way to force thinking about what our future might look like. But it’s also time-consuming. So wouldn’t it be easier if I could hand off a good chunk of it to my trusty Claude? I figured we should try, at least after I’d established a bit of a style/format. The wait-until-there-are-examples plan follows what we had success with when drafting emails for financial advisors, and matches everything I’d tell you about context. It’s even consistent with what you’d do when hiring a ghostwriter. So after relegating codebrain (my thought-partner repo that I run with both Claude and GPT models) to editor duties for the first 8 pieces, I decided to let it drive on the 9th post. I’d love to say it was planned out carefully, but the truth is I was just behind and stressed about some stuff and wanted to take it off my plate. I was surprised by what I found. First, the prompt. I told codebrain I wanted a post comparing a few systems of agent memory. My codebrain setup is the #1 thing my friends ask me about (since leaving the basic environments of the Claude or ChatGPT desktop apps can be intimidating), so I figured we could look at what I did along with a couple other memory-related systems in the AI ether that had some real juice. After I laid out this high-level vision, I asked it to give me three Fable-written drafts along with three from GPT-5.6 Sol. I’ve settled into using Fable in Claude Code as my go-to for most work-related tasks (which tend to be pretty code heavy) and the Codex desktop app with Sol as my daily for light fun/personal work (way better at browsing through film camera eBay listings or finding current Alaskan ferry schedules). So I’m familiar with both – but I’d never actually read them side-by-side. Now, the outputs. I was shocked. I’ve heard Dan Shipper at Every [talk about his preference for Sol in writing tasks](https://every.to/vibe-check/vibe-check-gpt-5-6-sol-is-our-favorite-model-to-collaborate-with). But everyone writes differently and I tend to be less bothered by Claude-isms than others (though tbh I’m starting to become a bit pickier). Besides, I’d given them plenty of content to pattern-match. I didn’t think they’d be that different. But Sol was night and day better (I preferred all 3 of Sol’s drafts to Fable’s). This held for [the next piece](/notes/the-wire-works-both-ways/) as well – and that time codebrain mixed the drafts up so my read was blind. I picked Sol again, and honestly I could tell who wrote what. Two openers for the exact same assignment:
“In last week’s piece I said the context you owe your agent comes in three pieces — prompts, memory, and examples of what good looks like — and then spent the whole post on the third. Today: memory. Instead of just walking through my setup, I want to lay out the three most interesting personal memory systems I know of. They showed up within two months of each other this spring, built from the same parts, disagreeing completely about how to use them. By the end you should be able to build one — or hand this piece to your agent and have it build yours.”
“In my last piece I called memory one of the three big pieces of context you can give an agent. Then I moved on after roughly a paragraph. Helpful! So let’s actually build the thing.”
first is fable, second is sol
Ok, so I preferred Sol. What else surprised me? I was startled by how eerie it was reading Sol’s caricature of my voice (and the ick I got from Fable’s Claude-ized version of me). I also felt a little smug having successfully taken the first step towards taking some work off my plate. And about that – while it shifted the work, I actually think I spent longer reading and editing than I would have if I’d just written it in the first place. (The Sol draft [did ship](/notes/make-it-remember/), by the way – edited, disclosed, Sol’s name on the byline.) Since I’m trying to figure out exactly where I stand and what the process should look like, I’ll keep revisiting rather than treating this as the conclusion to an AI-writing experiment. This is very much a work in progress. But let me leave you with a few thoughts/suggestions in case you want to start handing off more writing to our jagged and quirky agentic friends: Not groundbreaking, but think about what you’re writing and why. Internal docs for alignment or knowledge transfer are very different from voice-first thinkboi pieces, and should probably be treated differently (I’ve found you can get pretty organized, concise writing from any model as long as you prompt it to do so). Try different models. Seriously. As an Anthropic fanboy, I was surprised by how much I prefer Sol’s writing. I plan on doing more comparisons to see if I can optimize further. One caveat: the preference doesn’t carry over to working with these models – I’m happy to deal with Fable’s quirks because I do enjoy something about how we think together on most things (I find Sol a little dry and sycophantic, comparatively). If style matters, give examples. I’m guessing neither model needed 8 essays to figure out that I prefer choppy, imperfect writing, but I’m also curious to see how things evolve as we continue to grow this corpus. Maybe Fable can close the gap (though I doubt it). Solving for voice doesn’t necessarily solve for thinking. While Sol does a pretty good impression of me now, I’ve been very disappointed at how it misses what I think is important sometimes. It still doesn’t think like me. A good parrot, but far from a clone. Obvious but worth saying – you’re still responsible for what you write. That’s less a lesson from this experiment and more something that matters in professional settings. Prompting alone can get you clear, structured, dry writing. It can’t do the crafting for you. Very excited to revisit this after getting more reps in. And geez, it’s exhausting thinking about how much of the future’s writing will be AI’s (and probably done with a lot less care). See you in slopland!