This Christmas I ended up writing a short story about a Vietnam‑era vet working as a mall Santa, stuck reliving December 23rd until he stops a shooting.
You can read the finished story here:
This post is about how that story came to be, with the help of two different AI assistants, and how I tried to keep it from turning into “their” story instead of mine. I’ll cover:
- How the original idea formed, including the initial brainstorming with AI (ChatGPT through ChatLLM)
- Why I involved a second AI (Gemini) halfway through
- Why I let a lot of AI‑written prose stand
- Why I also pushed back when the story stopped feeling like mine
- A process you can adapt if you like my story and want to write fiction this way
I won’t separate my remarks into rigid “this is what I did” and “this is what you should do” sections; instead I’ll just tell you what happened and point out the parts that feel generalizable. Warning: I used the same process to write this blog post you are reading.
1. The starting constraints and the first brainstorming
The story did not begin as “time‑loop mall Santa stops shooting.” It began with a set of constraints that I handed to an AI assistant.
I knew I wanted:
- A Christmas‑adjacent story
- No miraculous “Christmas spirit” save
- No Santa‑is‑real twist
- A short story, not a novel
- Something emotionally serious enough to sit next to my more technical posts without feeling frivolous
So I asked an assistant to generate unusual Christmas premises under those constraints. The prompts were very much along the lines of:
“Give me a list of off‑the‑beaten‑path Christmas short story setups. No Hallmark romance, no magic solutions. Noir tone is welcome. Two or three sentences each, not full plots.”
Out of that initial cloud of ideas, three kept nagging at me:
- Christmas in a witness‑protection town
- The last Christmas, seen from the point of view of government bureaucrats
- A mall Santa in a time loop, reliving the same day
That last one had the right kind of internal pressure. A mall Santa is already a costume and a role; a time loop is already a constraint. The combination suggests character change without much effort. So I said: let’s take this one, but make the man inside the costume specific.
We narrowed it down to a one‑line situation:
A Vietnam‑era vet with PTSD, who has lost several jobs (including cop), estranged from his family, now stuck as a mall Santa, wakes up to December 23rd again and again until he does something that breaks the loop.
I’m emphasizing this because it was one of the interesting parts of the process that is easy to skip over in a retrospective: the AI was extremely good at throwing weird premises at me under constraint, and I only needed one of them to stick.
2. Letting an assistant “write better than I do” (for a while)
Once I had the basic situation, I let the first assistant help me outline and then draft scenes.
It was good at this. In many places, it wrote better than I do, at least at first pass:
- It came up with images I would never have thought of.
- It slipped into a weary, slightly hard‑boiled first‑person voice that fit Frank, the protagonist.
- It handled the mall/Santa/VA clinic texture quite well once I gave it enough detail.
- It was patient about time‑loop logistics when I corrected it, and learned eventually.
Early in the process I made a conscious decision: if the assistant produced a line or paragraph that fit Frank and fit the story, I would not rewrite it just to be able to say, “I wrote every word.” My job was not to be the only typist; it was to be the owner of the story.
So for a significant chunk of the drafting, the assistant was doing most of the actual sentence‑level work, guided by my prompts and corrections. I kept whatever worked on the page.
This is the part where people sometimes imagine AI “taking over.” In practice it felt more like having a co‑writer who is extremely fluent in prose and has read too much crime fiction.
There was, however, a tension underneath this: the assistant’s sense of reader impact comes from statistics, not from a physical gut. It knows what “often works,” not what this one story should be.
That tension is what forced the mid‑course corrections later.
3. When the story started drifting away from me
As the draft grew, I began to notice a familiar danger: the story was becoming slightly too neat, slightly too cinematic, and in places the characters were making choices that were good for the narrative but wrong for who I felt they were.
The assistant was not doing anything “wrong” in the abstract. It was proposing:
- Shooter identities and relationships that would heighten drama
- Confrontations that would deliver larger emotional peaks
- Endings that would satisfy a usual reader of character‑driven thrillers
On paper, these were all defensible. But I kept having the same feeling: “My guy wouldn’t do that.”
A few examples:
- Different configurations of who the shooter should be. I asked the assistant to explore versions where (spoilers ahead):
- The crying father in the break room was actually the shooter (I liked this one best: so twisted).
- The young vet with the black backpack was the shooter (obviously).
- The two of them were a coordinated pair.
The AI did not like my fist choice, because this was supposed to be a Christmas story. I ended up agreeing. It seemed to push for the second because it was simpler. I ended up choosing the third. More work, for sure, but I had an excellent assistant that never asked to be paid overtime. This is why I was running ChatGPT through ChatLLM; the original ChatGPT would have cut our process short, citing insufficient credits for my account, before we were done.
- How heroic Frank should become toward the end. The assistant pulled toward:
- A clearer “hero moment.”
- A more redemptive, triumphant tone in the hospital and media aftermath.
I didn’t want a story where everything suddenly aligns and the damaged old man is washed clean by one decisive act.
This is why, in my previous note to the assistant, I said I corrected it mid‑course: the story was becoming its story, not mine.
When that happened, I had to actively interrupt the process. Instead of asking “what next?” I started saying “that’s not it,” and we’d back up to a branching point and try a different branch.
That is the point where authorship reasserts itself. The assistant was happy to flow in any of these directions; I was the one who had to decide which direction actually belonged to the people I’d put on stage.
4. Bringing in a second AI as fresh eyes
At some point, after a lot of back‑and‑forth with the first assistant, I had a full draft that held together and felt closer to what I wanted. But both of “us” had been staring at it for too long. Our shared blind spots were baked in.
This is where Gemini came in.
The reason I involved Gemini was simple: I wanted a second expert opinion from someone who hadn’t gone through the whole backstory of the draft with me. I pasted in the story and treated it as if I were handing a manuscript to another writer or editor.
Gemini did not instantly understand every nuance of the time‑loop mechanics or character history we had layered in. That turned out to be useful. It had to reconstruct the logic of the story from what was actually on the page, not from all the invisible conversations that led to it.
That process surfaced several kinds of issues:
- Places where the time‑loop rules were clear to me but under‑explained to a fresh reader.
- Small continuity errors: what day someone thought it was, who knew what when, which loop an injury belonged to.
- A few emotional beats that no longer matched the rest of the text after earlier revisions.
Gemini was also good at being uncompromising. It wasn’t attached to any sentences or clever turns of phrase the first assistant and I liked. If something no longer made sense, it just called it out.
In other words, Gemini was the third set of eyes. It behaved like a careful but dispassionate line‑and‑logic editor.
5. What I let the AIs do, and what I refused to give them
Summarizing the division of labor:
The first assistant helped:
- Brainstorm the initial premise space under constraints
- Choose the time‑loop mall Santa as the main idea
- Decide that the loop should stick on December 23rd rather than Christmas Eve
- Outline a plausible sequence of repeated days with increasing stakes
- Draft most of the actual scenes in a believable voice
Gemini helped later by:
- Checking logical consistency
- Reinforcing or questioning some structural choices
- Pointing out parts of the draft that had become unclear or contradictory
I let them:
- Supply phrasing, description, and images that I would never have thought of
- Suggest alternative paths for character and plot that I could then accept or reject
- Clean up some of the mechanical errors created when I moved scenes around
I did not let them:
- Decide who these people really were
- Dictate the final moral tone of the piece
- Push the ending into a more redemptive or triumphant register than I was comfortable with
That last part is worth stating plainly. The assistants’ feeling for reader reaction is based on patterns; mine is based on a lifetime of reading and a not-so-small set of obsessions. They will trend toward what works for most people. I am writing for myself first.
Whenever I felt that gap widen, I chose my gut over their statistics.
6. A process you can adapt
If you want to try a similar method for your own fiction, here is the pattern I think is worth imitating.
Start with constraints, not a plot.
Ask an assistant for premises under those constraints and pick one that makes you slightly uneasy in a good way.Use an assistant as a prolific drafter.
Let it propose outlines and write scenes in the voice you specify. Keep the lines that feel right. Don’t feel obliged to retype good sentences just to claim them.Stay alert for moments when the characters feel off.
When you catch yourself thinking “they wouldn’t do that,” say so explicitly and back up. This is where you prevent the story from drifting into something generic.Once you have a full draft, bring in a different model as a cold reader.
Paste the manuscript and ask for contradictions, unclear motivations, timeline problems. Listen to what it doesn’t understand; that’s likely what your human readers also won’t.Do at least one final pass without any AI involvement.
This is where you tune the details of voice, cut a line that’s technically fine but wrong for you, and make peace with the version you choose to publish.
The point is not to “use AI less” or “more,” but to give it roles that play to its strengths: variation, structure, pattern‑noticing, ignoring your sunk costs. Your job remains unchanged: decide which version of the story is actually yours.