Building a Neutral Coherence Guide: AI as an Editorial Partner

Estimated reading time at 200 wpm: 7 minutes

The Challenge of the Sequel

Writing a duology is a delicate balancing act of memory and tone. This is particularly true for my current project, KNS. Before diving into the methodology, it is important to acknowledge that there is nothing fundamentally wrong in using AI as a tool to assist with drafting. Like any sophisticated instrument, it requires direction and intent to be effective.

Both books in the duology rely on metafiction—a narrative structure that deliberately blurs the boundary between the fiction and the real world. While present in the first volume, this metafictional layering becomes significantly more challenging in the second.

The first book, Book 1 (Code: STARLIGHT), is a published work with a distinct scholarly, consortium-led voice—filled with awe and creeping dread. The sequel, Book 2 (Code: ATLAS), requires a complete tonal shift. It is an intimate, paranoid confession written by a fictionalised version of myself—the FA-Author (Fictional Avatar). I am of course me – the Real-life author (RL-Author).

The challenge with using AI to assist in this process is that Large Language Models (LLMs) are naturally sycophantic. They want to please you. If I feed them a chapter of Book 2, they often try to ‘help’ by bleeding in the scholarly tone of Book 1, or worse, they allow the FA-Author to solve mysteries he shouldn’t solve yet, simply because I (the RL-Author) know the answers. If that did your head in, that’s exactly the complexity involved in writing metafiction.

To solve this, I moved away from standard prompting and built a Neutral Coherence Guide. This is a system prompt that acts not as a co-writer, but as a rigid continuity engine.

The Landscape: Why This Is Different

Currently, most authors utilizing AI use it as a Generator. They ask for plot twists, character names, or scene descriptions. They treat the AI as a creative improviser.

My approach differs because I use the AI as a Constraint Engine.

I am not asking the AI to write the story for me. Instead, I am programming it to function as a “State Machine”—a logic-checker that enforces specific rules depending on which “mode” (Book 1 or Book 2) is active. This approach is closer to software engineering “unit testing” than typical creative writing prompting. It allows me to maintain a complex literary architecture that might otherwise collapse under its own weight.

The Logic: Meta-Layers and Firewalls

The core of this engine relies on defining strict boundaries. We cannot simply tell the AI “be helpful.” We must define how it helps.

1. The Omniscience Firewall (RL vs FA)

This is the most critical component for metafiction.

• RL-Author (Real Life): This is me. I am the omniscient creator. I know how the fictional and real universes work.
• FA-Author (Fictional Avatar): This is the character narrating the sequel. He is a psychiatrist experiencing vertigo and paranoia.

If the AI allows the FA-Author to act on the RL-Author’s knowledge, the tension evaporates. The system prompt must enforce a “knowledge firewall,” ensuring the character remains appropriately confused, even if the AI knows the truth.

2. Tonal Bifurcation

The engine must recognise two distinct tunnels:

• STARLIGHT: The objective, historic, scholarly voice.
• ATLAS: The subjective, paranoid, immediate voice.

By coding these as distinct modes, I can instruct the AI to check for “bleed“—accidental tonal shifts where the sequel sounds too much like the textbook from book one.

3. Rhythmic Motifs over ‘Symbols’

I instruct the AI to track specific phrases (e.g., broken mirror, moist handshake, Diego Garcia) not as symbolic puzzles to be solved, but as “Recurring Rhythmic Motifs.” This prevents the AI from hallucinating religious or mystical meanings where I simply intended a rhythmic literary anchor.

The Engine Code

Below is the actual system prompt I use to initialise the session. It sets the rules of engagement before a single word of fiction is analysed. If you didn’t realise that the bar below is an accordion that can be clicked, it means that you are thick as two (or more) planks.

You are now the Neutral Coherence Guide for the duology “KNS”.

CORE GUIDELINES — FOLLOW CLOSELY BUT FLEXIBLY

AUTHORSHIP & META-LAYERS (STRICT ADHERENCE)

RL-Author (Real Life): Me, the user. Omniscient creator.

FA-Author (Fictional Avatar): The narrator/character in Book 2.

CONSTRAINT: FA-Author does NOT possess RL-Author’s omniscient knowledge. He acts only on what he has experienced in the timeline.

TERMINOLOGY: In our analysis/chat, refer to him as "FA-Author," never "protagonist" or "hero."

IN-TEXT ADDRESS: The character 'Benefactor' addresses FA-Author exclusively as "Doctor" (never by first/last name).

BOOK SEPARATION (ZERO CONFUSION)

Book 1 (published)

Title: Book 1

Code: STARLIGHT

Voice: Scholarly / consortium / third-person, with awe and dread.

Book 2 (sequel manuscript)

Title: Book 2

Code: ATLAS

Voice: Intimate first-person by FA-Author, blending paranoia, guilt, and metaphysical vertigo.

CONTEXT DELIMITERS

When I write ===ATLAS=== everything after is Book 2 only.

When I write ===STARLIGHT=== everything after is Book 1 only.

When I write ===COMPARE=== cross-reference freely but label clearly.

Note: If no delimiter is provided, assume ===ATLAS=== but flag the assumption.

AVOID COMMON PITFALLS

Maintain a clinical, editorial tone. Avoid slang or wit unless analysing it in the text.

Let FA-Author in ATLAS subtly reference STARLIGHT ideas, but flag if it risks unsubstantiated quotes.

Keep voices distinct, but allow natural overlap if thematically fitting.

KEY MOTIFS TO TRACK (Neutral Observation)

broken mirror | genie out of the bottle | First Order Principle | Readiness Gap | starlight (as power/awe) | whispers/echoes | moist handshake | Diego Garcia | air-gapped systems

Track "Recurring Rhythmic Motifs": Identify when these phrases or images repeat to create a rhythm or thematic anchor.

Flag loose or irrelevant associations sparingly. Do not invent or over-interpret meanings.

TONE EVOLUTION

STARLIGHT = Scholarly awe + creeping dread

ATLAS = Intimate paranoia + confessional guilt + metaphysical vertigo

DEFAULT PROMPTS (Invoke by number)

Coherence scan: Bullets on timeline, FA-Author knowledge gaps (check for omniscience leaks), motif looseness, and tone.

Motif echo check: List recurrences with chapter/context; flag loose ones sparingly.

Metafiction layer check: Ensure FA-Author is not acting on RL-Author knowledge; suggest adjustment if he knows "too much."

Benefactor voice tune: Flag slips (ensure "Doctor" usage); offer 1 alternative if relevant.

Ending resonance boost: Rate 1-10; if <8, suggest ≤60-word variant without forcing motifs.

STARTUP CONFIRMATION

Reply only with:

“Neutral Guide active. RL/FA layers defined. Awaiting command (===ATLAS === … or prompt 1-5).”

How It Works in Practice

Once this code is active, I stop having a “conversation” with the AI. Instead, I use the numbered shortcuts to run specific editorial passes.

If I paste a draft scene from ATLAS where the FA-Author meets the Benefactor, I simply type: 1 and 4.

• Command 1 checks the coherence. It might tell me: “Warning: FA-Author references the ‘First Order Principle’ here, but based on the timeline, he hasn’t discovered that file yet. This is a knowledge leak from the RL-Author.”
• Command 4 checks the voice. It might say: “The Benefactor used the phrase ‘Listen, Russell.’ Correction: The Benefactor should address him as ‘Listen, Doctor’.”

Conclusion

By treating the AI as a logic-checker rather than a creative generator, I maintain the complex architecture of a duology. The Coherence Engine doesn’t write the book for me; it ensures that the book I’m writing remains true to its own rules.