AI Character Consistency & Plot Hole Detection

Language models have no persistent memory. Every generation call starts from zero. The AI that wrote chapter one and the AI that writes chapter twenty-three are, functionally, strangers who happen to share a vocabulary. Without intervention, the results are predictable and painful.

Eye color changes mid-book. A character described as an only child in chapter four gains a sister in chapter sixteen. The antagonist's motivation shifts from revenge to greed and back again without explanation. The protagonist's fear of water, established in an early flashback, is never mentioned again — even when the climax takes place on a boat. The tone of a character's dialogue drifts from sharp and sardonic to warm and earnest because the model defaulted to a more generic voice.

These aren't edge cases. They're the default outcome of generating a long manuscript with AI unless every generation call carries enough context to keep the model honest. Entangled Text treats consistency as an engineering problem and solves it with structured data: character bibles, contextual injection, negative prompts, name blacklists, and outline-driven generation. Each layer catches a different category of drift.

A character bible is a structured reference document for each character in your book. In Entangled Text, the character bible field is injected directly into every chapter generation call. The AI doesn't have to remember who your characters are — it's told, explicitly, every single time.

A good character bible entry covers the facts that matter for consistency: physical description, personality traits, speech patterns, key relationships, backstory details that affect behavior, and any constraints the character operates under. The more specific the entry, the harder it is for the AI to drift.

This entry travels with every chapter. When the AI writes chapter nineteen, it knows the scar is on the left arm, that Maren doesn't swear, and that she can't arrest anyone. Consistency isn't maintained by memory — it's maintained by repetition.

Character bibles handle who. Context-aware generation handles what happened. Every chapter generation call in Entangled Text includes three layers of narrative context that keep the AI grounded in the story as it exists, not the story it might invent from scratch.

• The full outline — The AI sees where the story is going. It knows the betrayal happens in chapter fourteen, so it plants the right seeds in chapter twelve. It knows the subplot resolves in chapter twenty, so it doesn't accidentally wrap it up in chapter fifteen.
• Previous chapter summary — A compressed version of what just happened. This gives the AI immediate continuity: who was in the scene, what emotional state the protagonist was in, what was said, and what was left unresolved.
• Recent content window — The tail end of the previous chapter's actual prose. This anchors tone, voice, and pacing so the transition between chapters feels seamless rather than jarring.

Together, these three inputs mean the AI is never generating blind. It knows the characters, it knows the plan, it knows what just happened, and it knows how the last scene sounded. The result is AI manuscript feedback that reads like a continuous narrative rather than a series of disconnected writing samples.

Telling the AI what to write is only half the problem. The other half is telling it what not to write. Language models have strong defaults — certain phrases, patterns, and narrative moves they gravitate toward unless explicitly blocked. Negative prompts act as guardrails that prevent specific categories of inconsistency and bad habits.

A negative prompt is an instruction that says "do not do this." It gets injected alongside the character bible and context window, and the AI treats it as a hard constraint for the entire generation call.

Negative prompts are especially powerful for recurring problems. If the AI keeps defaulting to a cliché you hate, adding it to the negative prompt list kills it permanently. If a character trait keeps slipping — the AI makes your stoic protagonist suddenly emotional — a negative prompt locks it down. Think of these as the rules your AI editor is not allowed to break.

One of the most common AI consistency failures is name collision. The model invents a new character in chapter nine with a name suspiciously close to an existing one. Or it introduces a "Sarah" in a scene where a different "Sara" already exists. Or it recycles a name from a character who died three chapters ago, creating confusion about whether the reader is supposed to recognize them.

Entangled Text's name blacklist prevents this. You define which names are reserved for your established cast, and the AI is blocked from introducing new characters with those names or close variants. If your protagonist is named Elena, the AI won't create an "Elana" or "Helena" as a throwaway shopkeeper in a market scene.

The blacklist also catches unintentional name drift — the subtle kind where the AI starts spelling "Katharine" as "Katherine" or shortens "Benjamin" to "Ben" when the character is always called by his full name. These small inconsistencies compound across a long manuscript and erode the reader's trust in the narrative. The blacklist enforces naming discipline the AI would otherwise lack.

Forgotten subplots are the plot hole most readers notice first. A mystery is raised in chapter three and never answered. A promise between characters is made and ignored. A ticking-clock deadline passes without comment. These aren't failures of creativity — they're failures of tracking, and they happen because the AI has no mechanism for remembering what it set up earlier unless you give it one.

The chapter outline serves as that mechanism. Each entry in the outline contains plot beats — the specific events, reveals, and turning points that must happen in that chapter. When the AI generates chapter twelve, it reads the outline and sees that this is where the stolen letter resurfaces, where the protagonist confronts the ally she suspects of lying, and where the B-plot takes a sharp turn. None of these can be skipped because they're in the spec.

This structure also prevents the opposite problem: premature resolution. Without an outline, the AI tends to wrap things up too early. It resolves tension the moment it creates it because unresolved tension requires tracking state across multiple chapters, which stateless models can't do on their own. The outline forces patience. If the outline says the betrayal reveal happens in chapter twenty, the AI can't pull it forward to chapter thirteen just because the scene lends itself to drama.

Prevention is ideal, but sometimes the inconsistencies already exist. Maybe you generated chapters before setting up character bibles. Maybe you wrote the first draft yourself and want AI manuscript feedback to catch contradictions you missed. Maybe you inherited a manuscript from a collaborator and need to audit it for continuity.

Entangled Text's Edit a Book feature handles this. Paste your entire manuscript, write an editing instruction focused on consistency, and the AI processes it chunk by chunk with full context from surrounding chapters. You can target specific problems:

• Character audit — "Check every physical description of Maren across the full manuscript. Flag and fix any inconsistencies with her character bible."
• Timeline check — "Verify that all date references and time-of-day details are internally consistent. The story spans fourteen days starting on a Monday."
• Name standardization — "The character's name appears as both Katharine and Katherine. Standardize to Katharine everywhere."
• Subplot continuity — "The stolen documents subplot is introduced in chapter 4 but never resolved. Add resolution threads in chapters 16 and 19."

Every change appears as an inline diff. You review, accept, or reject each one. The AI writing consistency tool doesn't silently rewrite your book — it shows you exactly what it wants to change and waits for your approval.