Using AI to Write Fiction: How to Get Rid of That “Machine Flavor”

1. The Core Problem: Not a Weak Model, but an Overly Flat Workflow

When using large language models to generate narrative text, many developers encounter a familiar dilemma: the plot advances correctly and the information is clear, yet the writing still feels “off.” Specific symptoms include:

• Dialogue that is too polished: Characters speak as if taking turns reading from a script. Every line is complete, polite, and logically consistent, lacking the interruptions, omissions, rhetorical questions, and emotional spillover of real speech.
• Abrupt emotional shifts: One moment characters are arguing, the next they calmly explain background information, with no transitional action or psychological buffer.
• Formulaic environmental descriptions: The wind is “howling,” the night is “silent,” the rain is “torrential.” These are not wrong, but they lack any connection to the character’s state or the narrative tension.
• Homogenized character voices: You define five characters with different backgrounds, personalities, and experiences, yet they all speak as if they attended the same onboarding session. Remove the names, and it sounds like one person giving sequential reports.

These issues appear to be about writing quality on the surface, but they are fundamentally structural problems in the generation workflow. When a single prompt is responsible for planning, dialogue, narration, pacing, and more, the model tends to choose the “safest” expression—the most regular, template-like output. To escape the machine flavor, you cannot simply pile on more rules. Instead, you need to break the task into layers and let models or sessions better suited to each type of work handle specific problems.

2. Overall Approach: Layered Generation Instead of One-Shot Drafting

A practical direction for improvement is to split “writing the prose” into four stages: planning, script drafting, dialogue revision, and prose expansion. Each stage should have clear inputs, outputs, and acceptance criteria.

┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│  Planning   │ →  │   Script    │ →  │  Dialogue   │ →  │    Prose    │
│  Plot/      │    │   Lines/    │    │  Revision   │    │  Expansion  │
│  Conflict   │    │   Actions   │    │  Naturalness│    │  Narration/ │
└─────────────┘    └─────────────┘    └─────────────┘    └─────────────┘

This layering delivers two direct benefits:

1. Quality improvement: With the dialogue revision layer and script layer working together, characters finally know to “shout over the gunfire” or “yell for help as the carriage slides,” instead of standing at an accident scene giving a situation report.
2. Unexpected efficiency gains: Full prose generation time dropped from over 30 minutes to about 12 minutes. Even accounting for the script layer, the total still saved roughly 10 minutes. The reason is that the script layer locks in key actions and dialogue intent ahead of time, so the prose expansion model no longer needs to repeatedly “guess” what a character will say or do next.

3. Key Practice 1: Give Dialogue Its Own Revision Pass

3.1 Why Dialogue Needs Separate Handling

Generating full prose requires a relatively generous token budget, so it is unrealistic to route every word through the most expensive, quota-constrained model. A more practical approach is to let a lower-cost model handle the bulk of the prose, then send clearly problematic exchanges to a separate session for dedicated processing.

In early experiments, it was possible to have the same model revise itself, but in practice this often produced only minor word changes while the overall conversational pattern remained unchanged. For example, one character lists two or three pieces of information in order, another confirms each item, and a third summarizes—every line serves a purpose, but together it still reads like meeting minutes. Deleting a few repeated words does not end the meeting.

3.2 Key Points for Dialogue Revision Prompts

The goal of the dialogue revision layer is not “polishing” but changing the structure of the exchange. Prompts should include the following constraints:

• Allow incomplete sentences: Spoken language is full of omitted subjects, half-finished thoughts, and repeated emphasis.
• Allow interruptions and overlap: One character starts responding or objecting before another has finished speaking.
• Require responses to attitude, not just information: Characters should first react to the other person’s emotional stance, then process the information itself.
• Limit the frequency of “he said/she said” tags: Use actions, pauses, and interjections to replace some dialogue tags.
• Ban sequential report-style structure: Do not allow the fixed loop of A delivers complete information → B confirms → C summarizes.

A simplified revision prompt example:

Rewrite the following dialogue passage with these requirements:
1. Include at least two interruptions or overlapping lines;
2. Give each character at least one incomplete sentence;
3. Have characters respond to the other person's emotion before responding to the information;
4. Remove more than half of the "he said/she said" tags, replacing them with actions;
5. Do not allow a structure where three characters confirm the same information in sequence.

3.3 Acceptance Criteria

Revised dialogue can pass a simple test: If you cover the character names, can you still tell who is speaking? If the lines still sound like the same person giving sequential reports, the revision has not addressed the “character voice” level. In that case, return to the character profile cards and extract each character’s distinct verbal habits again.

4. Key Practice 2: Introduce a Script Layer to Pre-Organize Lines and Key Actions

4.1 What the Script Layer Solves

When the prose model receives only a plot outline, it must simultaneously decide what characters say, what they do, what the environment is like, and how fast the pacing should be. This kind of multi-objective optimization easily degenerates into template-based expression.

The script layer’s role is to pre-freeze the “what to say” and “what to do,” allowing the prose model to focus on “how to write it.” Specifically, this includes:

• The key action sequence for each scene: who moves first, who moves later, and what the physical consequences are.
• Each character’s dialogue intent: whether a line is meant to pressure, probe, conceal, or call for help.
• Emotional turning points: where the mood shifts from tension to relief, or from cooperation to confrontation.
• Information release order: what is said first, what is delayed, and what is deliberately withheld.

4.2 Output Format for the Script Layer

The script layer does not need complete sentences. It is closer to “action–dialogue intent” pairs:

Scene: The carriage begins to slide
- Character A (technical lead): Grabs the handrail, tries to determine if the brakes have failed → Dialogue intent: short, urgent command, no explanation
- Character B (observer): Sees a vehicle approaching from behind, voice half-sliced by the wind → Dialogue intent: shouting, but the information is incomplete
- Character C (medical lead): Thrown to one side, curses first, then asks "Is anyone still able to move?" → Dialogue intent: masking fear with emotion

With a script like this, the prose model only needs to “translate” actions and intent into narration and dialogue during expansion. It no longer needs to invent the conversational structure on its own. This is exactly where the efficiency gain comes from: separating planning from expression reduces the model’s trial-and-error during generation.

5. Key Practice 3: Model Assignment and Quota Allocation

5.1 Assign Models by Task Type

Different models show notably different performance across tasks. A pragmatic allocation strategy:

Task Type	Recommended Model Characteristics	Reason
Plot planning	Strong reasoning, long context	Needs to maintain global consistency
Script layer	Strong structured output	Needs to break plots into actions and intent
Dialogue revision	Strong colloquial feel, low “polish”	The more formal the model, the more likely it produces meeting minutes
Prose expansion	High throughput, controllable cost	Largest volume of text, needs sufficient supply

One notable observation: for dialogue revision, a lower-tier model may actually perform better because it is furthest from “that strange AI polish and tidiness.” Some sentences come out more complete, and some redundant explanations get removed—both are real gains.

5.2 Avoid “Another Guessing Round After the Draft”

If a planning model only says “this passage feels unnatural” after the draft is complete, then sends the original writer back to guess again, the process becomes extremely inefficient. A better approach is to let the planning model participate directly in the script layer, locking down the conversational structure before prose generation begins. That way, the prose model does not need to “guess”—it only needs to “write.”

6. Common Problems and Troubleshooting Checklist

In practice, the following issues recur frequently and can serve as a self-check list:

Dialogue still reads like meeting minutes

• Check whether the script layer only captured “information delivery” without noting “emotional reactions.”
• Check whether the dialogue revision instructions allow interruptions, omissions, and incomplete sentences.
• Check whether each character’s verbal habits are clearly distinguished (word choice, sentence length, tone).

Emotional transitions feel abrupt

• Check whether the script layer marks action buffers before and after turning points.
• Check whether the prose expansion was instructed to “write action first, then dialogue, then internal state.”

Environmental descriptions feel formulaic

• Check whether environmental details are tied to the character’s current goal. Wind, rain, and light are only worth describing when they affect a character’s actions or emotions.
• Avoid standalone environment description paragraphs. Embed the environment within action wherever possible.

Character voices sound the same

• Create a “verbal habit card” for each character: common phrases, sentence length, whether they interrupt others, whether they use rhetorical questions, whether they avoid direct answers.
• Enforce the “names covered, still distinguishable” requirement in the dialogue revision layer.

7. Summary

The key to removing the “machine flavor” from AI-assisted writing is not switching to a stronger model but changing the structure of the generation workflow. Making dialogue revision its own layer, introducing a script layer to pre-organize lines and actions, and assigning models by task type can significantly improve text naturalness while keeping costs under control—and may even deliver an unexpected boost in generation efficiency.

This method is not tied to any specific platform or tool. Any text generation environment that supports multi-session, multi-model collaboration can apply it. The essential point is this: do not let a single model handle planning, dialogue, narration, and pacing control all in one generation pass. Once you layer the process so each stage focuses on one job, the machine flavor naturally fades.