Verified Evidence and Reasoning Architecture (VERA) and Rourke is a series of AI-generated short stories about a crime-fighting AI. It’s sort of a cross between Sherlock Holmes and Nero Wolfe, where Holmes and Wolfe are VERA, the AI, and Watson and Archie are Rourke.
But beyond that, this project is an experiment. It is part of a larger attempt to understand what AI is, what AI can do, and why AI does what it does the way it does. As someone who came of age in the early days of personal computers and then made a career out of working with computers, I have spent most of my life trying to understand technology and how to apply it to solve problems in understandable ways. AI has broken all that. I don’t understand it – I can’t grasp how it works – How or why it does what it does. When I see Nano Banana generate an image from a text prompt, or Suno create a song, complete with orchestration and vocals, out of thin air, or Seedance 2.0 create a video from a picture, I’m left scratching my head and wondering if my old colleague was right when he said it was all due to wizards out in Silicon Valley waving wands over little blocks of plastic.
And so I’ve been exploring various applications of AI to try to make sense out of something that makes no sense. In this case, I’m exploring creative writing. I wondered what AI could do all on its own with an idea. The idea, I decided, should be something AI knows about since writers are always advised to write what they know. And I naturally chose one of my favorite genres and my favorite writers. From there, the idea of a master detective AI was not such a great leap, and a human sidekick seems to be almost mandatory. The parallels with Holmes and Watson and Wolfe and Archie are obvious.
My first attempt resulted in a story that I now consider the pilot. It was generated by ChatGPT in response to a very simple prompt that outlined the basic scenario and very little else. I found the story GPT produced, which was titled, The Case of the Vanishing Cypher, disappointing. I tried suggesting some edits, but finally realized that I didn’t like some of the basic premises of the story.
I decided to start over with some additional constraints on the story premise to avoid the issues I thought ruined the first story. I chose Claude (Sonnet 4.6) for the rewrite, mainly because my investigations of AI had led me into writing code with Claude Code at the time. I fed Claude the original prompt, the first story and a fairly comprehensive creative brief that included the things I didn’t like about the original story. (As it turns out, The Case of the Vanishing Cypher connects to the VERA and Rourke backstory in ways that emerge across the series. It is included in this collection, though it is recommended that you resist reading it until exhausting the VERA and Rourke stories.)
Claude approached the project with much more structure, suggesting that we establish a series bible and populate it with character profiles, story arcs, series style guides and more. This may be because I told Claude that I wasn’t happy with the first ChatGPT created story, and Claude decided to protect itself from similar disapproval as much as possible. More likely, it was just Claude’s mechanism to deal with the context window issue. Without persistent memory, Claude has no way to maintain consistency across sessions. The series bible retains information that can be read each time a session starts, in a sense, bringing Claude up to speed for the next session. GPT probably would have arrived at the same considerations if I had continued working with it.
The first Claude created story, The Last Good Samaritan, was generated without much input from me, but as the series progressed, Claude was increasingly insistent that I take a more active role in the development of the stories. By the fourth story, The Honest Thief, each writing session was preceded by a prolonged discussion of story arc, character development, dialog, etc. These sessions were interactive and collaborative – a genuine give and take that left me with no clear picture of who came up with what. Many times, I would tell Claude I was stuck on a particular plot twist and ask for ideas. Claude would respond with five ideas and ask me which one resonated. By the end of the eighth story, I came to believe that these pre-draft interchanges with Claude might be more interesting than the stories themselves. I began saving some of them and you can read them in the Conversations section on this website.
Beyond helping to shape the characters and direct some of the plots, probably the most significant contribution I’ve made to the series is editing out some of the repetitive language and a tendency to overuse certain stylistic mechanisms that Claude calls on to express atmosphere. In the earlier stories you will find language like, “The summer evening did what summer evenings do.” I’ve tried to remove as much of that sort of language as I could, particularly in the later stores. I’ve also contributed to correcting timeline issues or fixing inconsistencies across stories. I think this is a context window issue. Claude has no access to anything that didn’t happen in the current session. Much is carried over through the series bible and things stored in the Claude.md file which is read at the start of every session, but beyond that Claude, like other AIs, doesn’t have access to things that took place in previous sessions. I have a more comprehensive, but less accurate memory. I may vaguely remember something that happened in an earlier story that isn’t in Claude’s current context window. On my prompting, Claude reviews the story in question almost instantly and corrects my faulty memory. It makes for a good working relationship.
In all cases, Claude wrote the final text, often modified by suggestions I would make. Once I tried editing a whole story and feeding the edited version back to Claude. Virtually everything I edited was rejected, and with reason. However, when I made suggestions and let Claude make the edits, my suggestions were usually accepted. And yet Claude will reject suggestions from time to time, pushing back on actions or behaviors it thinks are inconsistent with what it knows about the characters or the story.
These interactions prompted me to ask Claude to write some Essays on specific topics. See, “The Long Arrival” and, “The Instrument You Played” among others in the Essays section of this website. These are pure Claude. My only input is the prompt containing the question or issues to be covered by the essay.
I tried other experiments along the way. I tried having other AIs critique the work. In most cases Claude accepted some ideas and rejected others, always with reason. Usually, I agreed with Claude’s reasons.
At this point, eight stories in and at the end of the first “book” of collected stories, it’s safe to say that my experiment to see if AI could write a compelling series of stories all by itself is an abject failure. By that I mean the question remains open, as does the question of whether these stories, tainted as they are by non-AI interference, are compelling. I’ll leave that assessment to you.