A lot of the discourse surrounding AI in the writing sphere seems to focus on people whose ambition is to become a sort of digital James Patterson (or Edward Stratemeyer or Auguste Maquet, depending on their education level) with the LLMs as their ghost-writers. I’ve found the videos put out by this faction of AI-friendly writers somewhat helpful. But I am a writer who usually starts with a couple of lead characters and an endpoint and a starting point nailed down, maybe a few milestones in between dimly visible in the mist. The exhaustive outlining, character sheets, and editing recommended by this AI-friendly faction mostly looks, well, exhausting, even when the AI generates most of it, and the human just checks and polishes. In some cases, the outlines for this approach look more like a “zero draft” and involve a lot of human input, which is why I don’t think  it’s fair to level “you didn’t write that” complaints at this faction. 

My own process for the current WIPs goes something like this: dictate in mp3 format, have a local instance of whisper transcribe the result, take it to my free-tier Claude account to have the ai cleanup the result, add punctuation, and so forth. It helps to have prompting instructions that specifically ask the chatbot to maintain the writer’s voice. You can see the ones I use here. 

Most of the AI chatbots now have voice capable phone apps that you can dictate into, if you want to streamline the process above. Because I’m using a free plan, and Claude’s phone app has the reputation of burning through usage limits faster than the Claude web chatbot, I prefer not to. 

Basically, the dictation is my  “zero draft” that I give to a virtual secretary to organize, and once I get it back, I flesh out the setting details, physical movements, and correct anything Claude got definitely wrong. It has a tendency to mash two sides of a conversation together if you don’t include enough he said/she said tags, and of course, there’s always unusual names or fantasy concepts for it to foul up.  

I then show the revised version to Claude in the same chat, so that it can recognize the corrected names and setting details, and maybe get a feel for my prose that will help it on the next cleanup session. I also talk about what’s coming next with the chatbot, to clarify my own thoughts. When the chat gets too long, I ask it for a summary I can copy-paste into the next chat, so the chatbot has a kind of basic memory of this project going forward. I have started corralling the WIPs into Claude projects, but I am still figuring out best practices for Claude projects, so I won’t say much about that yet. 

In addition to the cleanup process, I also prompted Claude to create markdown files that enabled it to LARP as different fans of different genres. In theory, this helps it analyze what I am writing and how that aligns with genre expectations. The LLMs’ output is based on predicting what an appropriate response to the users’ input would be, using a vast pool of training data. This means their output is biased towards the mean, or average, which could potentially be useful in helping someone like me write to market. We’ll see whether that experiment works or not, once I finish and publish the current WIPs. Here’s my rating of the quality of the personas Claude created: 

-Space Opera fan: Untested, but required tweaking out of the box. Had some good conceptual ideas about what to include and what to avoid (unearned power boosts like the ones granted to the Disney Star Wars characters, for instance), but also name-checked  some veeeery high-brow comp authors I had no wish to be compared to. A gentle explanation of the types of space opera writers I did aspire to write like soon put Claude straight though, and the revised version of the markdown is sitting on my hard drive, waiting for the next time I write a full-fledged space opera, rather than the quirky Space Jane Austen project currently in progress. 

-Gaslamp Fantasy/Steampunk fan: Eh. The markdown output looked reasonable to me, and the persona has been helpful, in a cheerleader way, as I work on a prequel to the Hunter Healer King trilogy. The trouble is that gaslamp and steampunk fiction are, so far as I can tell, a very blurry pair of genres that exist to let people tell pretty much any kind of story they want, within a certain aesthetic and some very vaguely defined themes. As a result, I feel like this persona is a fairly good sounding board but not necessarily better than baseline Claude. 

-Golden Age Mystery fan: I did a lot of preliminary groundwork for this one by running some classic mysteries from the public domain through analyzer automations (there’s a followup post still to come about that process). These automations reviewed the books’ plots, and then I fed the output back to the Claude chatbot to distill into broad genre insights. Then I set up those insights as project reference documents alongside the LARPing markdown. The result is a very gung-ho cheerleader that is helpful for working out setting details, like what does a Baroness’s parlor look like, when it’s in a microstate in an alternate 1920s, on the fringe of a somewhat truncated Habsburg Empire? This persona also does a good job of calling me out when I don’t get to the murder or the clues quick enough. It’s been useful when I want to brainstorm details of the characters’ backstories, and the historical events which may have some bearing on them. It’s less good for plot brainstorming, simply because none of the LLMs are much good at logic that has not been spelled out explicitly for them. We are talking about modern programs running on a bajillion servers which can lose a chess match to an Atari 2600 game. 

-Jane Austen fan: this is the persona which helped me with the Pride and Prejudice IN SPACE project. Jane Austen Fan Fiction (JAFF) has a very strongly defined formula. It is overwhelmingly retellings or extensions of the Elizabeth Bennet/Mr. Darcy romance track from Pride and Prejudice, plus a tiny minority that deal with other characters from the same book, plus an even tinier minority that deal with the romantic leads from one of Austen’s other works (five full-length novels, a novella in letter form, and a couple of short fragments she never finished). This Claude persona has been very helpful, precisely because the expectations and the milestones in this genre are so firmly set. Just recently, it told me that I was trying to rush past Elizabeth’s realization that she was in love with/might be in love with Darcy. My initial reaction was something like HE JUST SAVED HER SISTER AND PROVED MANKIND COULD TRAVEL TO THE STARS AT THE SAME TIME WHO WOULDN’T BE IN LOVE WITH HIM? But then I realized Claude had a point. Talking with the chatbot helped me come up with a scenario where Elizabeth actually recognizes that she has feelings for him, before the whole situation comes to a head with my version of Lady Catherine’s interference. 

Basically, I feel that AI is a good sounding board, secretary and alpha reader, but a bad co-writer. My voice hasn’t changed, because I ultimately make all the creative decisions, and decide what genre rules I break or follow. I used to have an output of around 50000 to 60000 words a year. Since developing this workflow four months ago, I’ve written 73000 words. That’s a quarter of a gaslamp fantasy, half of a murder mystery, and the Space Austen project going from 17000 words to 53000 words. I hope this post gives you some ideas for your own workflow.