Usually, I hold the “making of” posts for these fanfics until a bit later in the serialization, but I wanted to report on this one while my thoughts were fresh.

A Curious Case started out, the way all these AI-drafted fanfics do, with a story idea that I wanted to read but didn’t feel passionate enough to actually write. I knew it was going to be a long, complicated story to tell. I also discovered that Anthropic was running a promotion through most of June and the beginning of July that doubled the user’s session limits when working inside Claude CoWork.

Now, CoWork is kind of an odd beast, basically the Chat function but with the ability to work in folders. It used to be its own tab inside Claude Desktop, but recently it became one half of a toggle (with the Chat option) that appears under the initial message box when you open Claude. Supposedly it works in the web and mobile apps too, but I have not tested this.

A lot of users dismiss CoWork in favor of Claude Code (which has its own tab), but I’ve been too annoyed by Code’s demands that I get a GitHub account to pay much attention to Code. I had dabbled with CoWork a little bit already. It reformatted a Gutenberg text for me once, outlined (but did not write) Sense and Sensibility and Placage. It also made some cover mockups for me using images I’d generated in Google’s Nano Banana, and a piece of ad art (based on the existing book cover) for Pride & Planetoids. A Curious Case would be the largest project I had undertaken in CoWork.

I knew from past experience that the Sonnet 4 line of models did not play well with CoWork. Possibly Sonnet 5 has changed that, but I haven’t tested it in CoWork yet. At the point in time we’re talking about, at the tail end of May or the start of June 2026, Opus was the only game in town for CoWork.

Rather than blasting through my account limits, I sat down with Sonnet 4.6 in chat, where I had done some preliminary brainstorming on the concept for A Curious Case, and it helped me set up a plan for generating A Curious Case in CoWork: a folder for the project, with subfolders labeled Reference, Output and Approved. Reference held the following:

• a txt copy of Pride and Prejudice downloaded from Gutenberg
• a series bible including a “writer’s room” of authors I wanted Claude to try to emulate
• a continuity tracker for Claude to update after reviewing the files I had approved
• a subfolder for individual scene briefs

Everything except the Gutenberg file was in markdown (*.md), which is some kind of text file that the AIs seem to favor.

I had Sonnet draft the “scene briefs” inside a Claude Project with guidelines for fanfic, based on the open source “Narrative Physics Engine” concept. I would read the scene briefs, which summarized what needed to happen, what needed not to happen, that kind of thing, and then ask Sonnet to revise. When I was happy with the result, I would move the scene brief to the Reference folder. To allow room for flexibility, one scene at a time would be drafted in CoWork. I would review it, make changes (in VSCode), and move to Approved and get Claude to update the continuity tracker.

Fable 5, Anthropic’s latest frontier model, showed up when I was in the early stages of messing around with this. Most of the early Elizabeth chapters were generated by Fable, I think down to around 8-9 inclusive. Then Fable went away, and most of the rest of the remaining Elizabeth chapters were written by Opus 4.8. Then Fable came back, and the last five or eight Elizabeth chapters were also written by Fable, over the sweltering Independence Day weekend when my AC failed and I had no brainpower for my own writing projects. Fable helped plot out the Bingley chapters as well, but due to me running out of Fable usage, Opus 4.8 wrote them.

Of the two models, Fable was somewhat better at dialogue and mood and made a more amusing sounding board. But it also went more spectacularly off the rails when it got anything wrong. Opus 4.8 was adequate: it didn’t have Fable’s highs, but it also didn’t have Fable’s lows. It tended to work more quickly and cheaply, which was helpful when I told it to correct a plot point or excessive exposition. When asked to review approved chapters to update the continuity tracker, both models invariably homed in on the one or two typos I had made when revising the latest chapters. VSCode came in handy here because they would generally refer to which line in VSCode the typo appeared on, which made the typos easier to track down. I did not distinguish myself as a continuity editor on this project: you will note that Charlotte invites Elizabeth to Kent in two consecutive chapters. My bad.

The writers I wanted Claude to try to emulate were Georgette Heyer, Dorothy Sayers, and Margery Allingham. A good chunk of Sayers is public domain, including classics like Clouds of Witness and Strong Poison. Allingham’s and Heyer’s works are still mostly under copyright, and I knew from past experience that Claude was only slightly familiar with Heyer’s more famous works. However, Allingham’s Crime at Black Dudley and Heyer’s The Education of Philip Jettan are both public domain. They are also fairly characteristic works in terms of style and tone, even if they aren’t the best things either author ever wrote. I figured they at least would be in the training data. I don’t know if I would say that either of the Claudes really achieved that feel, but they did succeed in sounding somewhat early twentieth-century British to me.

I finished by having Opus in CoWork convert my edited/approved versions of the markdowns back into a Word document and then scheduling the result on WordPress. If I had to do this over again, I think I would try for a more even balance between Bingley and Elizabeth as narrators. I’ve also discovered the spaCy Python library, which can be used to analyze things like word frequency, sentence and paragraph length, etc., in public domain authors. It might be worth seeing if I can use that information to get the AIs to draft in a more natural way. I’ve seen other people demo the idea. For now, it’s clear that I personally can’t get Claude to write fiction worth paying money for, so the claudefics continue to be free content on my website, generated for my amusement and hopefully yours as well.