Two doctors, both named Adrian
Imagine your novel is set in a remote sanatorium. There are two doctors there: Dr. Adrian Vossberg, the cool head physician with the blue eyes, and Dr. Adrian Vosskuhl, his warm-hearted young assistant with the brown eyes. To you, these are two completely different people. To an AI that has your entire book in view all at once, they're two very similar-sounding names — and before you know it, one of them ends up with the other's eye color.
Mistakes like these plague AI writing tools constantly, as soon as a book grows large. We took a hard look at the problem — and something surprising happened along the way. An honest report from the workshop.
The problem: the AI gets everything at once
When Hermes writes a paragraph for you, the AI needs context: who's in the scene, where it takes place, what the tone is. We assemble that context from your book — from your characters, places, plot, and style.
Until now, we did it with a sledgehammer: we handed the AI everything on every request. All the characters, all the places, every note. For a short story with five people, that works just fine. But in a book with thirty, eighty, a hundred entries, the three or four characters who actually matter in this scene drown in a sea of the irrelevant.
Two mistakes then happen especially often:
- Mix-ups. The two Adrians. Or two women named Marlene — one the head nurse, one the patient. The more similar names in the book, the greater the risk that the AI swaps their traits.
- Accidental spoilers. Because the AI sees your whole structure, it might reveal in chapter 2 that a character believed dead is actually alive — because the summary of chapter 5 gave it away.
The idea: a codex that shows only what's needed
Authors who know tools like NovelCrafter or Sudowrite will recognize the principle under names like "codex" or "story bible": the AI doesn't get everything, only the entries that belong to the current scene. Like a good assistant who hands you exactly the right file — not the whole archive.
Sounds simple. It isn't.
The honest part: our first attempt made everything worse
Before we ship something like this to you, we test it. So we built our own test book — a small, dense story with traps deliberately built in: the two Adrians, the two Marlenes, a brother believed dead who only turns up later. Then we had the same chapters written once before and once after the change, and compared which version made fewer mistakes.
The result of the first version was sobering: it didn't get better, it got worse.
The reason was instructive. Our first, naive selection searched the text for full names. The catch: in scene instructions, the heroine is simply called "Elise," not "Elise Vanderhoek." So the main character fell out of her own scenes. At the same time, a shared surname pulled in too much — mention "Elise Vanderhoek" and along came the dead brother and an uncle of the same name who was never even present.
The good part: the test caught the mistake before it ever reached you. That's exactly what testing is for.
The fix — and what came out of it
The second version chooses more cleverly. It includes a character if their full name appears — or a name fragment that belongs to only one single character (a "Strup" is unambiguous, a "Marlene" is not). It understands German inflection ("Elises suspicion"). And the main character can be pinned in place, so she never gets lost.
The result, measured cleanly across several runs:
- Fewer name mix-ups, fewer spoilers — and no category got worse.
- The biggest single gain right where things had been most stuck before: the chapter with the two Marlenes.
And then we tested it even bigger. We built a Hanseatic merchant saga in the tone of Buddenbrooks — five families, 119 entries, and, as in real families, seven Wilhelms, six Heinrichs, and five Annas spread across rival houses. A nightmare for "everything at once."
The result was unambiguous: the bigger the world, the bigger the gain. In this large book, the AI gets roughly 96% less irrelevant clutter per request — and still keeps hold of the right characters. With three women named Anna at the same ball, it reliably picked exactly the right three, not the maids who happened to share the name.
What this means for you
- Your characters stay themselves — even in a three-hundred-page book with a large ensemble.
- Fewer accidental spoilers, because the AI no longer has your entire structure in front of it.
- And the AI no longer wastes its "attention" on a hundred irrelevant entries — more of it stays for your scene.
Honest stays honest: this isn't magic. We make sure the right information is in the context — we can't force the model to use it perfectly in every sentence. And it works best when you give your characters distinguishable names. Two "Adrians" stay harder than an Adrian and a Bernhard.
What's next
That was the first step: the right context at the right time. Next, we want to highlight name mentions directly in the editor and make them clickable, enable "spoiler-safe" character states up to a given chapter, and — for series — a book-spanning bible.
If you're curious how Hermes handles long texts and coherence in general, read also Context management: how AI keeps track and Story coherence: characters who stay true to themselves.
Give it a try
Write your book with a tool that thinks along — and with a team that looks closely before anything reaches you. Start free with Hermes 3000.