How to Make AI Write in Your Style: A Fiction Writer's Guide to LLM Settings

You gave the AI your plot outline. It wrote a chapter. The plot is fine. The pacing is fine. But the voice is wrong — it reads like a Wikipedia article cosplaying as a novel. Generic adjectives, predictable sentence structures, emotional beats that feel earned by no one.

The AI isn't bad at writing. It's bad at writing like you.

The default output of any LLM is an average of its training data — competent, inoffensive, and utterly without personality. Making it produce prose with a specific voice requires understanding three layers of customization that most writers never touch. This guide covers all three, with concrete examples you can use in any AI tool.

Layer 1: System Prompts (The Foundation)

The system prompt is your most powerful tool. It tells the AI who it is and how it should behave before it sees any of your actual content.

What Most People Write

You are a creative writing assistant. Help me write my novel.

This tells the AI almost nothing. It will produce generic fiction.

What Actually Works

You are a fiction writer specializing in noir detective stories set in 
1920s Shanghai. Your prose style is:
- Short, punchy sentences. Rarely exceeding 15 words.
- Heavy use of sensory details — smells, textures, sounds.
- Dialogue is clipped and subtext-heavy. Characters rarely say what 
  they mean directly.
- Avoid purple prose. No "orbs" for eyes, no "crimson liquid" for 
  blood.
- Internal monologue is sardonic and self-deprecating.
- Action scenes use short paragraphs, one beat per paragraph.
- Emotional moments are shown through physical reactions, not stated 
  feelings. "Her hand shook" not "She felt afraid."

Why this works: Specificity constrains the AI's output space. Instead of choosing from all possible writing styles, it now has concrete rules to follow. The more specific your rules, the more distinctive the output.

System Prompt Template for Fiction

You are writing a [GENRE] novel. The tone is [TONE].

Prose style rules:
- Sentence length: [short/medium/long/varied]
- POV: [first/third limited/third omniscient]
- Tense: [past/present]
- Dialogue style: [naturalistic/stylized/sparse/verbose]
- Description style: [sensory-heavy/minimal/poetic/clinical]
- Emotional expression: [show don't tell/internal monologue/both]

Avoid:
- [specific clichés to avoid]
- [specific words or phrases to avoid]
- [specific patterns to avoid]

The writing should feel like: [reference author or book for tone]

Reference authors are powerful shorthand. "Write in the style of Cormac McCarthy" gives the AI a rich model to draw from — sparse punctuation, Biblical cadence, unflinching violence described matter-of-factly. You don't need to describe all of this explicitly; the reference does the work.

But don't only use a reference. Add your own specific rules to push the output away from a generic imitation toward something that's influenced by, but not copying, the reference.

Layer 2: Temperature and Sampling Parameters

Temperature controls how "creative" vs "predictable" the AI's word choices are. Most writers never change this. They should.

What Temperature Actually Does

Temperature 0.0-0.3 — Conservative. The AI picks the most probable next word almost every time. Prose is grammatically perfect but predictable and flat.
Temperature 0.4-0.7 — Balanced. The AI occasionally picks less obvious words, producing more varied and interesting prose while staying coherent.
Temperature 0.8-1.0 — Creative. More surprising word choices, unusual metaphors, unexpected sentence structures. Also higher chance of nonsense.
Temperature 1.0+ — Experimental. Increasingly chaotic. Occasionally brilliant, often unusable.

Recommended Settings by Genre

| Genre | Temperature | Why | |-------|------------|-----| | Literary fiction | 0.7-0.9 | You want surprising metaphors and unusual phrasing | | Thriller/Mystery | 0.4-0.6 | Clarity and pacing matter more than poetic language | | Romance | 0.6-0.8 | Emotional language benefits from some creativity | | Hard sci-fi | 0.3-0.5 | Technical accuracy and precision are key | | Fantasy | 0.6-0.8 | Worldbuilding descriptions benefit from creativity | | Horror | 0.7-0.9 | Unsettling prose needs unexpected word choices |

Top-P (Nucleus Sampling)

If your tool exposes it, Top-P is another lever. It controls the pool of words the AI considers:

Top-P 0.9 — considers the top 90% most probable words (default, works well)
Top-P 0.5 — only considers the top 50%, producing more focused but less surprising output

Practical advice: Start with temperature 0.7 and Top-P 0.9. Generate a chapter. If the prose feels too safe, increase temperature by 0.1. If it feels too random, decrease by 0.1. Small adjustments make a noticeable difference.

Layer 3: Persona and Context Instructions

This is the most underused layer. Beyond the system prompt, you can provide additional context that shapes the AI's output in ways the system prompt can't.

Novel-Level Instructions

These apply to your entire novel:

This novel takes place in a world where electricity was never 
discovered. Technology advanced through steam and clockwork. 
Characters should never reference or imply electrical technology.

The protagonist narrates unreliably — she omits details that make her 
look bad and embellishes her successes. The reader should gradually 
realize she's not trustworthy.

These instructions prevent the AI from breaking your world rules or character conceit, even when individual chapter prompts don't mention them.

Chapter-Level Instructions

These apply to a specific chapter:

This chapter is a flashback to 10 years before the main story. 
The prose should feel warmer and more nostalgic than the present-day 
chapters. Use longer sentences and softer imagery.

Key constraint: the reader knows this relationship ends badly, but the 
characters don't. Write the happy moments with an undercurrent of 
dramatic irony.

Few-Shot Examples

The most powerful technique most fiction writers don't use. Give the AI a sample of your actual writing and tell it: "Write in this style."

Here is an example of the prose style I want:

"Maren counted the ceiling tiles again. Fourteen. Same as yesterday, 
same as the day before. The guard's footsteps came and went like 
clockwork — heavy boots on stone, always the same rhythm. She'd 
stopped flinching at them a week ago. Progress, she supposed."

Match this style: short declarative sentences, dry observations, 
minimal emotional language, details that reveal character state 
without stating it.

Why few-shot works so well: Instead of describing your style abstractly (which the AI may interpret differently than you intend), you're giving it a concrete target to match. It's the difference between telling someone "paint something blue" and showing them the exact shade.

Putting It All Together

Here's how the three layers stack:

┌──────────────────────────────────┐
│ Layer 1: System Prompt           │  ← WHO the AI is
│ Genre, tone, style rules,        │     (set once per novel)
│ reference authors, avoid list    │
├──────────────────────────────────┤
│ Layer 2: Temperature / Top-P     │  ← HOW creative
│ Controls word choice randomness  │     (set once, fine-tune)
├──────────────────────────────────┤
│ Layer 3: Context Instructions    │  ← WHAT specifically
│ Novel rules, chapter mood,       │     (updated per chapter)
│ few-shot examples                │
└──────────────────────────────────┘

Layer 1 sets the baseline identity. Layer 2 controls the creativity dial. Layer 3 provides the specific context for each generation.

Most writers only use Layer 1 (and poorly). Adding Layers 2 and 3 transforms AI output from "generic fiction" to "fiction that sounds like it was written by a specific author with a specific vision."

Tool-Specific Implementation

ChatGPT / Claude (manual)

System prompt: Paste at the start of every new conversation (or use Custom Instructions / Projects)
Temperature: Not directly adjustable in the chat interface. Use the API for control.
Context: Paste novel-level and chapter-level instructions manually each time.

API (OpenAI, Anthropic, etc.)

System prompt: system message parameter
Temperature: Direct parameter (0.0-2.0)
Context: Include in user or system messages

Noveble

All three layers are built into the novel settings:

System Instructions: Set once per novel, automatically included in every generation
Temperature / Top-P: Adjustable per novel in the model settings panel
Persona Messages: Additional per-novel context injected alongside character profiles and chapter context

The advantage of an integrated tool is that you set these once and they apply consistently across every chapter — no re-pasting, no forgetting to include your style rules on chapter 23.

The Style Iteration Loop

Getting the perfect voice usually takes 3-5 iterations:

Generate a chapter with your initial settings
Identify what's wrong — too formal? Too wordy? Wrong emotional tone?
Add a specific rule to the system prompt to address it: "Never use semicolons" or "Limit paragraphs to 4 sentences maximum"
Regenerate and compare
Repeat until the output consistently matches your vision

Save your final settings. They're now your novel's "voice configuration" — reusable, consistent, and independent of any single chapter.

Voice configuration is step one. For the complete workflow from idea to finished novel, see our complete guide to writing a novel with AI. And if you're just starting out, check out the 7 mistakes beginners make — skipping voice configuration is #6.

Want to configure all three layers in one place? Noveble gives you system instructions, temperature control, and persona messages — all applied automatically to every chapter. Set your voice once, write consistently forever. Free to start.