How to Prompt AI So Your Novel Doesn't Read Like a Machine Wrote It

Anyone who's spent more than a few hours using AI to assist with novel writing knows the feeling. You read back the output and something is off. The prose is competent, sure. Grammatically flawless. But it reads like a Wikipedia article wearing a trench coat and pretending to be a novel. The sentences are too clean, the metaphors too safe, the dialogue too articulate, and every character sounds like they graduated summa cum laude from the same MFA program.

The good news is that most of these problems aren't inherent limitations of AI. They're defaults — patterns the model falls into when you don't tell it otherwise. With the right prompting strategies, you can break those defaults and produce prose that reads like a human actually wrote it. This article walks through the most common tells of AI-generated fiction and gives you concrete prompts and techniques to eliminate them.

The Problem With "Good" Writing

Before we get into specifics, it helps to understand why AI defaults to the kind of prose it does. Large language models are trained on enormous datasets that include published novels, articles, essays, fan fiction, and everything in between. When you ask one to write fiction, it produces a statistical average of all that training data. The result is prose that's technically proficient but utterly generic — the literary equivalent of a stock photo.

Human writing, by contrast, is full of imperfections. Real authors have tics, habits, blind spots, and stylistic quirks that make their work distinctive. Cormac McCarthy didn't use quotation marks. Elmore Leonard's characters spoke in sentence fragments. Brandon Sanderson's action scenes read like choreography notes. These aren't bugs. They're what make those authors sound like themselves.

AI doesn't have a "self" to sound like, so it defaults to sounding like everyone and no one. Your job as a prompter is to give it a self — or at least enough constraints and instructions that it stops writing like a committee.

Tell #1: Purple Prose and Over-Description

This is probably the single most recognizable marker of AI-generated fiction. Ask an AI to describe a sunset and you'll get something like:

The crimson orb descended beneath the horizon, painting the sky in breathtaking hues of amber and violet, casting long shadows across the undulating landscape as the last vestiges of daylight surrendered to the encroaching darkness.

Every noun has an adjective. Every verb is doing too much work. Nothing is just red — it's crimson. Nothing sets — it descends. The sky isn't orange — it's painted in breathtaking hues of amber. This is prose that's trying to impress a creative writing teacher, and it's exhausting to read at novel length.

Real novels, especially commercially successful ones, tend toward cleaner prose. Here's how you might describe the same sunset in a progression fantasy:

The sun dropped behind the ridgeline. Kael pulled his cloak tighter and kept walking.

That's it. Two sentences. The reader knows it's getting dark and cold. The character is focused on something other than the scenery. Move on.

Prompts to Fix It

Add instructions like these to your system prompt or writing guidelines:

Prompt Example 1:

"Write in clean, direct prose. Avoid purple prose, flowery language, and over-description. Describe settings only when they're relevant to what the character is doing or feeling. Use concrete, specific details rather than vague atmospheric language. If a detail doesn't serve the scene, cut it."

Prompt Example 2:

"Follow the principle of 'one strong detail beats three weak ones.' When describing a location, pick the single most vivid or relevant detail and let it do the work. Don't stack adjectives. Don't describe things the POV character wouldn't notice or care about."

Prompt Example 3:

"Treat description like seasoning, not the main course. A few well-chosen sensory details grounded in the character's perspective are more effective than sweeping panoramic descriptions. The character is living in this world, not visiting it as a tourist."

Tell #2: Dialogue That Sounds Like an Essay

AI-generated dialogue is often the most obvious giveaway. Characters speak in complete sentences, use correct grammar, never interrupt each other, never trail off mid-thought, and deliver exposition like they're reading from a textbook. Here's a typical example:

"I understand your concern, Marcus. The situation is indeed quite precarious. However, I believe that if we approach the problem methodically, we can devise a solution that addresses both the immediate threat and the underlying structural issues."

No human being talks like that. Not in casual conversation, not in a fantasy world, not ever. Real dialogue is messy, incomplete, and context-dependent. People talk over each other. They use contractions. They start sentences and abandon them. They say "yeah" and "look" and "I mean" as filler. They speak differently depending on who they're talking to.

Here's the same exchange written like actual humans:

"Marcus, this is bad." "I know." "So what do we do?" He rubbed his jaw. "We figure out the immediate problem first. The rest we deal with after." "That's not a plan." "It's what we've got."

Prompts to Fix It

Prompt Example 4:

"Write dialogue that sounds like real speech, not written prose. Characters should use contractions, sentence fragments, and informal grammar. They should interrupt each other, trail off, and sometimes say things that are unclear or incomplete. Dialogue should reveal character through speech patterns, not through articulate exposition."

Prompt Example 5:

"Each character should have a distinct voice. [Character A] speaks in short, clipped sentences and rarely explains himself. [Character B] rambles and over-explains when nervous. [Character C] uses slang and profanity casually. Never let two characters sound interchangeable."

Prompt Example 6:

"Avoid 'as you know, Bob' dialogue where characters explain things to each other that they would both already know. If exposition needs to be delivered through dialogue, find a natural reason — a new character who genuinely doesn't know, an argument about how to handle something, or a briefing scene where the information review serves a tactical purpose."

Prompt Example 7:

"Characters should not speak in perfect paragraphs. Most spoken lines should be one to two sentences. Long monologues should be broken up by action beats, interruptions, or the other person's reactions. When a character does speak at length, it should feel like a deliberate moment — a speech, a confession, an argument — not the default mode."

Tell #3: The Thesaurus Problem

AI loves synonyms. It will use "murmured," "uttered," "proclaimed," "declared," "articulated," and "vocalized" when "said" would work fine. It replaces simple words with complex ones not because the complex word is more precise, but because it's different from what it used in the previous paragraph.

This extends beyond dialogue tags. Eyes don't look — they "gaze," "peer," "observe," or "scrutinize." People don't walk — they "traverse," "navigate," "make their way," or "proceed." The effect is prose that feels like it was run through a thesaurus filter, and readers pick up on it immediately.

Prompts to Fix It

Prompt Example 8:

"Use 'said' for the vast majority of dialogue tags. Avoid fancy alternatives like 'exclaimed,' 'declared,' 'murmured,' 'breathed,' 'intoned,' or 'opined.' When a tag other than 'said' is genuinely needed, use it sparingly. Often the best choice is no dialogue tag at all, replaced by an action beat."

Prompt Example 9:

"Prefer simple, common words over their more elaborate synonyms. People walk, run, look, and say things. They don't traverse, hasten, observe, or utter things. Use the fancier word only when it adds genuine precision — 'whispered' is fine when someone is actually whispering, but 'murmured' as a default replacement for 'said' is not."

Tell #4: Emotional Telegraphing

AI doesn't trust the reader. It will show a character slamming their fist on a table, then immediately tell you they're angry. It will describe tears running down a face and then add "she felt a deep, overwhelming sadness." The emotional state gets double or triple-communicated, first through action, then through internal narration, and sometimes through the reactions of other characters too.

Here's a typical example:

Kael's hands trembled as he stared at the ruined village. A cold fury burned in his chest, an icy rage that threatened to consume him entirely. He felt the anger rising within him like a tide, unstoppable and all-encompassing. His jaw clenched so hard his teeth ached, and he knew in that moment that whoever had done this would pay dearly.

That's four different ways of saying "he was angry." Pick one and move on.

Kael stared at the ruined village. His hands were shaking. He turned south and started walking.

The reader understands the emotion from context and action. They don't need to be told.

Prompts to Fix It

Prompt Example 10:

"Show emotion through action, body language, and behavior — not through internal narration that labels the feeling. Never write 'he felt angry' or 'sadness washed over her.' Instead, show what the character does with that emotion. Trust the reader to interpret. If you've shown it well, naming it is redundant."

Prompt Example 11:

"Avoid emotional redundancy. If a character's action already communicates their emotional state, do not follow it with internal narration restating the same emotion. One clear beat is stronger than three that say the same thing."

Tell #5: Formulaic Scene Structure

AI tends to write every scene with the same structure: establish setting, introduce conflict, characters discuss conflict, reach some resolution or cliffhanger, transition. Every scene gets roughly the same weight. Quiet moments get the same detailed treatment as action sequences. There's no variation in pacing, no scenes that are deliberately short or abrupt, no moments that linger longer than expected.

Real novels breathe. Some scenes are two pages. Some are twenty. Some chapters end mid-sentence. Some scenes are mostly dialogue with almost no description. Others are pure interiority with no dialogue at all. This variation in structure is part of what makes a novel feel alive.

Prompts to Fix It

Prompt Example 12:

"Vary scene length and structure. Not every scene needs a full beginning, middle, and end. Some scenes should start in the middle of action. Some should end abruptly without resolution. Short, punchy scenes can be just a few paragraphs. Let the emotional weight of the moment dictate the scene's length, not a formula."

Prompt Example 13:

"Match prose style to scene content. Action scenes should use short sentences, minimal description, and fast pacing. Quiet character moments can slow down with longer sentences and more interiority. Tense scenes should feel tense on the page — through rhythm, sentence length, and what's left unsaid."

Tell #6: Hedging and Qualifiers

AI prose is riddled with hedge words. Things are "somewhat" difficult. Characters feel "a sense of" unease rather than just feeling uneasy. Events are "rather" unusual. The sky is "almost" dark. These qualifiers drain energy from the prose and make everything feel tentative.

This habit comes from the model's training — it's been optimized to be accurate and measured, which is great for answering questions but terrible for fiction. Fiction needs commitment. If a character is scared, they're scared. Don't soften it.

Prompts to Fix It

Prompt Example 14:

"Eliminate hedge words and qualifiers. Never use 'somewhat,' 'rather,' 'quite,' 'a bit,' 'slightly,' 'almost,' 'seemed to,' 'appeared to,' or 'a sense of' unless the ambiguity is genuinely intentional and serves the narrative. Commit to concrete, definitive statements."

Prompt Example 15:

"Write with confidence and commitment. If something happens, it happens — don't soften it with qualifiers. If a character feels something, they feel it fully. Tentative prose reads as uncertain writing, not as nuance."

Tell #7: Repetitive Sentence Structures

AI tends to fall into rhythmic ruts. It'll produce three sentences in a row that all start with the subject, or a paragraph where every sentence is a compound sentence joined by "and" or "but." Read enough AI prose and you'll start hearing a metronomic quality — the sentences are all roughly the same length and follow the same grammatical pattern.

Good prose varies its sentence structure constantly. Short sentences punch. Long ones flow and build momentum before arriving somewhere unexpected. Fragments work. So do run-ons, when used deliberately. The goal is rhythm that serves the story, not a pattern that lulls the reader to sleep.

Prompts to Fix It

Prompt Example 16:

"Vary sentence structure and length deliberately. Mix short declarative sentences with longer, more complex ones. Use sentence fragments for emphasis. Avoid starting more than two consecutive sentences with the same word or grammatical structure. Read the prose aloud mentally — it should have a natural, varied rhythm."

Tell #8: The Banned Word List

Certain words and phrases appear in AI-generated text at rates far higher than in human writing. Readers and editors have started compiling lists, and seeing these words is an immediate red flag. Some of the most common offenders:

• "Tapestry" (as a metaphor for anything complex)
• "Delve" or "delved"
• "Intricate"
• "Nuanced"
• "Undulating"
• "Vestiges"
• "Palpable" (especially "the tension was palpable")
• "Myriad"
• "Labyrinthine"
• "Cacophony"
• "Ethereal"
• "Visceral"
• "Juxtaposition"
• "Enigmatic"
• "Whilst"
• "Amidst"
• "Amongst"
• "Interplay"
• "Reverberated"
• "Gossamer"
• "Iridescent"
• "Tendrils" (of anything: magic, darkness, fear)
• "Crimson" (used instead of red)
• "Azure" (used instead of blue)
• "Obsidian" (used instead of black)
• "A testament to"
• "Couldn't help but"
• "A dance of" (shadows, light, blades, anything)
• "Sent a shiver down his/her spine"
• "Let out a breath he/she didn't know he/she'd been holding"

None of these words are inherently bad. "Cacophony" is a perfectly fine word when you need it. The problem is frequency. AI uses these words constantly, and human readers have developed an almost unconscious sensitivity to them. When three or four show up in the same chapter, it reads as machine-generated even if every word was technically appropriate.

Prompts to Fix It

Prompt Example 17:

"Avoid the following overused words and phrases that are strongly associated with AI-generated writing: [insert your own list from above]. Use plain, direct alternatives instead. 'Red' instead of 'crimson.' 'Black' instead of 'obsidian.' 'Complex' instead of 'intricate.' If a word feels literary or fancy, it's probably the wrong choice — use the simpler version."

Prompt Example 18:

"Never use the phrase 'let out a breath he didn't know he'd been holding,' 'sent a shiver down her spine,' 'a dance of shadows,' 'a testament to,' or 'the tension was palpable.' These are cliché markers. Find a fresh way to express the same idea, or cut the line entirely."

Tell #9: Lack of Subtext

AI writes on the surface. When two characters are in conflict, they state their positions clearly. When someone is lying, the narration tends to flag it. When there's romantic tension, the internal monologue spells it out. There's very little subtext — very little of the meaning that lives between and beneath the words.

In real human fiction, some of the most powerful moments happen when characters say one thing and mean another, when the narration describes trivial actions while enormous emotional weight builds underneath, when the reader understands something the characters don't or won't say aloud.

Prompts to Fix It

Prompt Example 19:

"Write with subtext. Characters often say one thing while meaning another, especially in tense or emotionally charged moments. Not every thought needs to be spoken. Not every motivation needs to be stated. Let actions and dialogue choices imply what characters are feeling or thinking. The reader should sometimes have to read between the lines."

Prompt Example 20:

"Avoid having characters state their emotions or motivations directly in dialogue. People rarely say 'I'm angry because you betrayed me.' They say 'You know what, forget it' and walk out. They change the subject. They make a joke that isn't funny. They focus on an irrelevant detail because the real issue is too painful to address. Write dialogue that works on two levels — what's being said and what's actually happening."

Tell #10: Artificial Transitions and Connective Tissue

AI loves smooth transitions. Every paragraph flows logically into the next. Every scene change is preceded by a neat bridge sentence. Time skips are handled with graceful summary. The result is prose that feels frictionless in a way that actual novels rarely do.

Real fiction is comfortable with abrupt cuts. Chapters end on a line of dialogue. Scenes jump without warning. A character is in a tavern on one page and riding through a forest on the next with no explanation of how they got there. The reader fills in the gaps, and the story moves faster because of it.

Prompts to Fix It

Prompt Example 21:

"Don't over-explain transitions between scenes. You don't need to narrate a character traveling from point A to point B unless something important happens during the journey. Cut directly from one scene to the next. Trust the reader to follow the jump. White space between scenes is a perfectly good transition."

Prompt Example 22:

"Avoid summary paragraphs that bridge time gaps with language like 'Over the next few days...' or 'As the weeks passed...' Instead, cut directly to the next important moment. If the reader needs to know time has passed, a single concrete detail can establish it — changed weather, a healing wound, different clothes."

Putting It All Together: A Master Prompt

Here's a comprehensive prompt you can adapt and include in your system instructions when using AI for novel writing:

Master Prompt Example:

"You are assisting with a novel. Follow these writing rules strictly:

Prose Style: Write in clean, direct prose. No purple prose or over-description. Use concrete details, not atmospheric vagueness. Prefer simple words over fancy synonyms. 'Said' is the default dialogue tag. Eliminate hedge words like 'somewhat,' 'rather,' 'seemed to,' and 'a sense of.'

Dialogue: Characters speak like real people — fragments, contractions, interruptions, imperfect grammar. Each character has a distinct voice. No 'as you know, Bob' exposition. Most lines of dialogue are one to two sentences.

Emotion: Show, don't tell, and don't do both. One clear emotional beat per moment. No labeling feelings after showing them through action.

Structure: Vary scene length and pacing. Short sentences for action, longer ones for reflection. Not every scene needs a neat beginning, middle, and end. Cut transitions when possible.

Banned Patterns: Never use: tapestry (as metaphor), delve, palpable tension, vestiges, 'breath they didn't know they'd been holding,' 'a dance of [anything],' 'a testament to,' crimson/azure/obsidian as color words, tendrils of darkness/magic/fear. Avoid starting paragraphs with 'The' more than twice per page.

Subtext: Not everything is stated directly. Characters have unspoken motivations. Dialogue should work on multiple levels. Trust the reader.

Overall: If a sentence sounds like it could appear in any novel, rewrite it until it couldn't. Specificity and voice matter more than elegance."

A Final Note on Authenticity

The irony of this entire article is that you're using a machine to avoid sounding like a machine. That's fine. The goal isn't to pretend AI wasn't involved — it's to produce the best possible fiction regardless of the tools used. A carpenter isn't diminished by using a power saw instead of a hand saw, but they still need to know what good joinery looks like.

The prompts and techniques above aren't magic. They're guardrails. They keep the AI from defaulting to its worst habits, but they can't replace your taste, your instincts, or your understanding of what your story needs. You still have to read every line and ask yourself: does this sound like something a person would write? Does this sound like something my characters would say? Does this serve the story?

If the answer is no, rewrite it. Or re-prompt it. Or throw it out and write it yourself. The AI is a tool. You're the author. The difference between a novel that reads as AI-generated and one that doesn't isn't really about the AI at all — it's about how much you care about getting it right.