How to use Claude to co-write a historical epistolary novel (without it sounding like a Hallmark card)

Why the epistolary format is an LLM's secret superpower

Most AI-generated fiction is, frankly, unreadable. If you ask an LLM to write a standard narrative, it inevitably falls back on lazy tropes, breathless pacing, and an insufferable habit of wrapping up every paragraph with a neat little moral lesson. It acts like a high school student trying to hit a word count by throwing adjectives at the wall.

But epistolary writing—stories told entirely through letters, telegrams, diary entries, or logs—is different. The epistolary format relies on a single, distinct voice talking directly to another. It has built-in constraints, natural pauses, and a heavy reliance on subtext and voice. This plays perfectly to the strengths of Claude, which is arguably the most voice-capable LLM currently available.

Because the format relies on natural turns, you can treat Claude not as a black-box generator, but as a genuine writing partner. You write one letter; Claude writes the reply.

Here is how to set up an interactive, historical epistolary writing session with /platforms/claude that actually delivers gripping, high-tension drama instead of flowery nonsense.

Step 1: Establish the premise and the friction

Great letter-writing novels (think Dracula or The Screwtape Letters) do not feature two people politely updating each other on the weather. They thrive on conflict, dramatic irony, and secrets. Before you open a new chat, you need a premise built on friction.

For this guide, let us go with a classic Victorian rivalry: Two botanists in 1884. One is an established, slightly stuffy academic in London; the other is an eccentric, rogue collector in the cloud forests of Peru. They are hunting the same mythical carnivorous orchid.

To make this work, you must define the dynamic. LLMs default to being incredibly agreeable. If you write a letter accusing them of stealing your research, their default setting is to apologise and offer a cup of tea. You have to explicitly code conflict into the system prompt.

Step 2: The system prompt that bans the fluff

If you want to skip the generic "My dearest colleague" nonsense, you need a tight, opinionated prompt. Head over to our /prompts generator for structural inspiration, or use this specific framework in your initial Claude prompt:

`text We are going to co-write an epistolary novel set in 1884. We will write to each other in the form of letters, telegrams, and field notes.

You will play Dr. Alistair Finch, a stubborn, highly traditional London academic. I will play Evelyn Vance, a rogue botanist in Peru.

CRITICAL WRITING RULES: 1. NO melodrama. Do not use exclamation marks, breathless prose, or grand declarations of emotion. 2. Subtext is key. If Alistair is angry, he must express it through overly formal politeness, passive-aggressive professional snobbery, or pointed omissions. 3. Keep it historically grounded. Use 19th-century terminology, but do not make it sound like a caricature of Dickens. Keep the language dry, sharp, and precise. 4. Never summarize the plot or state the "theme" of our letters. Just write the letter. 5. Maximum length per letter: 350 words.

Please write Dr. Finch's opening letter to Evelyn, responding to her latest controversial paper published in the Royal Society Journal. `

Notice the strict constraints on word count and emotional expression. By limiting the model to 350 words, you force it to focus on voice rather than rambling exposition. You can learn more about how to manage these types of voice settings in our /glossary.

Step 3: Managing the back-and-forth

Once Claude replies with its first letter, the game is on. Your job is to push the story forward with your response, introducing new elements, lies, or discoveries that Dr. Finch has to react to.

Here is what makes this approach tick: you are not just directing the AI; you are playing a game of narrative chess. If Claude mentions that a specimen arrived in London completely rotted, you can reply with a letter hinting that you sent it that way on purpose to protect its secrets.

When writing your replies, keep these tips in mind: Inject sensory details:* Mention the dampness of the jungle, the smell of formaldehyde, or the ticking of a grandfather clock in the study. Use formatting:* Mix things up. Instead of a long letter, write a brief, urgent telegram. ("SAMPLES COMPROMISED. DO NOT OPEN THE CRATE.") Challenge the AI:* Introduce an unexpected plot twist in your letter and see how Claude's character rationalises it within their specific personality traits.

Step 4: Troubleshooting "agreeableness drift"

LLMs have short-term memory drift. After five or six exchanges, Claude might start to lose its sharp, Victorian academic bite and become soft, helpful, or repetitive.

If you notice Dr. Finch becoming a bit too friendly or modern, you do not need to restart the entire chat. Simply insert a brief course-correction in brackets before your next letter:

[System check: Dr. Finch is feeling increasingly threatened by Evelyn's success and should write this next letter with a much colder, more patronising tone. Maintain 1884 vocabulary.]

This immediate nudge resets the temperature of the conversation without breaking the flow of the document. If you hit persistent formatting or styling issues, check out the troubleshooting resources on the Claude support site.

The finish line

After a dozen exchanges, you will have a remarkably coherent, atmospheric, and highly original short story. Because you wrote half of it and guided the other half through constraint-based roleplay, it will feel genuinely collaborative.

When you are finished, copy the transcript, clean up the prompt headers, and you are left with a beautifully structured epistolary draft that has more character, grit, and tension than anything a generic "write a story about botanists" prompt could ever produce.