How I stopped the AI from fighting itself during session prep
December 10, 2025
Last week, I was trying to flesh out a new NPC and my AI assistant started doing that thing it does—the digital equivalent of tripping over its own feet.
It was supposed to be checking my notes for context. Instead, it searched for about three seconds, got bored, and started hallucinating a character that completely ignored the three months of lore I'd already established. Then, halfway through the description, it seemed to "wake up," realized it was wrong, and tried to course-correct mid-sentence.
The result was a total mess. A weird chimera of half-researched facts and generic, fantasy-generator filler.
I've seen this pattern enough to know what's happening: I was asking one agent to do too much. When you ask a single AI to search, write, and edit all in one go, you get the mushy middle. It's trying to be helpful in three different directions and ends up being mediocre at all of them.
So, I decided to stop asking it to be a generalist. I rebuilt my workflow into a "pipeline" of specialists: one that only searches, one that only writes, and one that only handles the filing.
It's a bit more "under the hood" than my usual posts, but if you're tired of your tools hallucinating your own campaign back to you, the payoff is worth it.
The fundamental issue is that searching and creative writing are different modes of thinking.
When you're searching, you want the AI to be a librarian: thorough, careful, and literal. You want it to notice that the Thieves' Guild contact is supposed to have a limp, or that the Harbor District is currently under a curfew.
When you're writing, you want a collaborator. You want something that riffs on ideas and finds a cool "voice" for the character.
Trying to do both at once is like trying to edit your prose while you're still drafting the first sentence. You lose the "flow state" of the writing and the accuracy of the research. In my old setup, I'd ask for a mysterious contact and the AI would get so excited about the "mysterious" part that it would forget the "contact" was supposed to be a double agent I'd already introduced.
The fix was to separate the jobs. I built a "Router"—basically a coordinator—that looks at my request and decides which specialist to call.
The Research Agent: This one is a librarian. It has no "creative" license. It scans my notes, finds the relevant threads, and hands back a structured brief: Here is what is canon, here is what's missing, and here are the constraints.
The Creative Agent: This one gets the brief and writes the prose. But here's the trick: it has no access to my search tools. It can't go hunting for more context. It has to work with what the researcher gave it. This constraint is actually a feature—it forces the research step to be better and keeps the writer focused.
The Edit Agent: This one is the clean-up crew. It can't write new stories; it just knows how to take a block of text and slot it into my document exactly where it belongs without breaking the formatting.
It sounds like more work, but it means I'm not spending twenty minutes fixing "hallucinated" lore.
One thing that surprised me: giving the AI a "perfect memory" actually makes it worse.
Early on, I gave every agent access to the full conversation history. Within ten minutes, the context window was filled with irrelevant tool outputs and old search results from three requests ago. It was digital clutter.
Now, only the Router has a memory. The specialists are stateless. They wake up, get a fresh brief for the specific task at hand, do the job, and then immediately "forget" everything. It's a clean slate every time. No accumulated baggage to get confused by.
Honestly? For some things, yes.
If you're just asking "What's the name of that tavern?" you don't need a three-agent pipeline. A single well-prompted bot is fine for that.
But if you have a campaign with hundreds of notes—where contradictions actually hurt the story—specialists are the only way to stay sane. It's about predictability. A single agent will drift; it will get "creative" with your facts. Specialists stay in their lane.
You probably aren't going to spend your Sunday afternoon coding a multi-agent router. But you can still use the logic.
Next time you're using ChatGPT or Claude for prep, don't ask it to do the whole job in one prompt. Break it up. Act as your own coordinator:
Step 1: "Based on my uploaded notes, give me a list of every NPC currently in the Harbor District and their current status."
Step 2: "Now, using only that list, suggest a new contact who wouldn't overlap with those roles."
Step 3: "Format that as a clean NPC entry for my records."
It's the difference between an exhausting hour of correcting the AI's mistakes and a refreshing twenty minutes of actually building your world.
That's the philosophy I've baked into Campaign Arks. It's not about having one "magic" button that writes your game; it's about having a set of specialized tools that handle the "work about work" so you can get back to the creative parts of being a GM.
If your AI output feels a bit mushy or generic lately, try narrowing its focus. Sometimes, the best way to get a better answer is to give the AI less to do.