The Averaging Problem
Here's something nobody tells you when you start feeding prompts to a language model: it doesn't obey instructions the way a human assistant would. A human hears "be firm, but don't be cold" and understands you want a specific emotional texture — warmth with a spine. A model often does something closer to math. It takes "firm" and "not cold" and finds a blended average, which usually lands somewhere in the neutral middle. Not firm. Not warm. Just... flat. This is the core problem with stacking positive and negative instructions in the same breath. You write "show her confidence, but don't make her arrogant," and the model doesn't hold both qualities in tension — it splits the difference until you get a character who's neither confident nor humble, just vaguely competent and forgettable. The negative constraint doesn't sharpen the positive one. It dilutes it. This happens constantly when people try to write a book with AI and pile constraint after constraint into a single paragraph of instructions, hoping the model will thread the needle. Sometimes it does. More often, you get prose that reads like it's apologizing for itself — hedged, cautious, drained of the specific choice you asked for.
A positive instruction tells the model where to go. A negative instruction, used carelessly, tells the model everywhere it can't go — which is a much bigger, blurrier space, and models default to the safest possible point inside it.
Diagnosing the Silent Cancellation
The sneakiest version of this problem doesn't show up in the first sentence. It shows up three sentences later, when the model, still "remembering" your negative constraint, undoes the positive instruction you gave at the top of the prompt. Say you ask for a scene where a character "finally stands up for himself" (positive) but also specify "don't make him aggressive or confrontational" (negative). Watch what tends to happen: sentence one, he stands up for himself, good. Sentence two, still fine. Sentence three, the model — anxious about violating the "don't be confrontational" rule — has him second-guess himself, soften his tone, or narrate his own hesitation. The scene you asked for quietly reverses itself before it even finishes. This is worth hunting for specifically when you edit a book with AI. Read your generated scenes and ask: does the emotional arc I requested actually survive to the last paragraph, or does it erode? A good habit is to isolate the exact sentence where the reversal happens. Nine times out of ten, it's triggered by the model trying to satisfy a negative constraint it interpreted too broadly. If you're doing a structured pass through a manuscript, this is exactly the kind of thing to flag during The Five-Pass Revision Order for AI-Assisted Novels — it's a distinct failure mode from continuity errors or pacing problems, and it needs its own read-through. There's a simple diagnostic test. Take any scene where the prose feels "off" but you can't say why, and ask yourself: did I give the model a constraint that contradicts, even subtly, the mood I asked for? Nine times out of ten, that's your answer.
The Layering Method
The fix isn't to abandon negative constraints — some of the most useful instructions you'll ever give are "don't do X." The fix is sequencing. Instead of dumping every positive and negative instruction into one undifferentiated paragraph, you layer them by priority, like you're briefing a director on a scene, not handing them a legal contract. Here's the structure that actually holds up:
- Layer one — the core positive instruction. One sentence, stated as an action, not a mood. Not "make it tense" but "he keeps his voice level even as his hand shakes."
- Layer two — the specific negative, scoped narrowly. Not "don't overdo it" but "do not have him explain why he's staying calm — no internal justification."
- Layer three — the positive replacement for what you just removed. This is the layer everyone skips, and it's the one that prevents the blandness. If you cut internal monologue, you have to tell the model what fills that space instead — physical detail, dialogue, silence.
Notice the order. Positive first, so the model commits to a direction before it learns what to avoid. Negative second, scoped tightly to one specific behavior, not a vague category. Positive replacement third, so the model isn't left in a vacuum where the safest move is to write nothing interesting at all. This sequencing matters more than most people expect. A model given "no adverbs" and nothing else will often produce sentences that are grammatically adverb-free but emotionally inert, because you've told it what to remove without telling it what carries the weight instead. That's the same failure pattern showing up again — negative constraints without positive replacements. If you're building out a story bible that AI models actually follow, this layering principle is worth baking into your character and scene instructions from the start, not just patching in during revision.
Worked Example: The Negotiation Scene
Let's build this properly, because the theory only means something once you see it in a real prompt. The scene: two rival trade envoys, formerly allies, now negotiating a treaty that both know favors one side. Our POV character, Maren, has every reason to blow up at the other envoy — he insulted her family's house in the previous chapter — but the scene calls for restraint. She holds it together. That's the positive instruction. The temptation for most writers is to add "but don't have her explain to herself why she's staying calm" as an afterthought, and that's exactly where the averaging problem creeps in — the model, trying to honor both, produces a version of Maren who's neither restrained nor raw, just numb. Here's a prompt that layers it correctly instead of dumping it flat:
Write the scene where Maren negotiates the grain tariff with Consul Birek. POSITIVE: Maren keeps her composure completely — she never raises her voice, never accuses him directly, and continues the negotiation as if the insult never happened. Show her restraint through controlled physical action: how she holds her pen, how long she waits before answering, the precision of her word choice. NEGATIVE (scoped): Do not include any internal narration where Maren tells herself to stay calm, reminds herself of her goals, or reflects on why she's suppressing her anger. No lines like "she reminded herself this wasn't about pride." REPLACEMENT: Instead of internal explanation, let her restraint show entirely through dialogue and micro-action — a pause before she responds, the exact temperature of her word choice, a detail she notices in the room that has nothing to do with her anger. The reader should feel the effort of her control without being told it's an effort.
Why this works: the negative is scoped to one specific technique — internal justification — instead of a vague mood like "don't make her seem too controlled." And the replacement clause gives the model somewhere to put the tension instead of just deleting it. Without that third layer, the model tends to solve "no internal monologue" by making Maren simply less angry, which flattens the scene. With it, the anger stays present, just relocated into gesture and pacing. If the first draft still hedges into blandness, that usually means the replacement layer wasn't concrete enough. Try tightening it further:
Revise this scene. Maren's restraint needs to read as costly, not effortless. Add one physical tell that recurs twice — for example, she straightens a single paper on the table each time Birek needles her, an unconscious gesture she doesn't notice herself making. Do not narrate her noticing it. Do not add any sentence that begins with "she felt" followed by an emotion word. Let Birek's dialogue get sharper across the scene while Maren's responses get shorter and more exact, so the contrast does the emotional work instead of narration.
This second pass is a good move any time restraint scenes come out limp — it's especially common in romance beat sheets where AI drafts break down during high-tension emotional scenes, and it shows up just as often in thriller interrogation scenes or fantasy court intrigue. The technique isn't genre-specific. It's about giving the model a physical outlet for suppressed emotion instead of an internal one.
A Reusable Template
Once you've done this a few times, you'll start noticing the pattern is repeatable across almost any scene where you want a character to do one thing while explicitly not doing another. Here's the template stripped down to its bones:
POSITIVE: [Character] does [specific observable action/behavior] in this scene.
NEGATIVE: Do not include [one specific, narrowly defined technique or line-type] — no lines like "[example of the exact thing to avoid]."
REPLACEMENT: Instead, convey this through [specific alternative technique — physical detail, dialogue subtext, pacing, a recurring object/gesture].
The three blanks that matter most are the specificity of the negative and the concreteness of the replacement. "Don't be melodramatic" is a category, and categories invite the model to guess wildly at what you mean, usually by removing more than you wanted. "Don't have her cry or raise her voice" is a behavior, and behaviors are easy to avoid precisely. Same goes for the replacement — "show it through actions" is still vague; "show it through what she does with her hands" gives the model an actual anchor. This template scales up nicely too. If you're plotting an entire chapter with several overlapping constraints — say, a magic system rule that can't be broken alongside a character voice that shouldn't slip into modern slang — stack multiple positive/negative/replacement trios rather than one giant paragraph trying to hold everyone accountable at once. This is the same instinct behind fantasy magic systems constraints AI will respect — rules work when they're itemized and scoped, not when they're buried in a wall of "also remember" clauses. It's also worth testing this template against your AI book outline before you draft full chapters. If a beat in your outline depends on a character holding back something important, write the positive/negative/replacement trio into the outline note itself, so every drafting pass for that scene starts from the same scoped instruction instead of you reinventing the constraint each time. One more thing worth knowing: different models handle this layering differently. Some hold the sequence better across long prompts than others, so if you're consistently fighting the averaging problem, it may be worth comparing performance across the best AI models for writing rather than assuming your prompt is the only variable. The documented differences in constraint-handling are real, and the Entangled Text documentation has more detail on model-specific quirks if you want to go deeper. If you're coming from a general chat interface, it's also worth reading how this compares in Entangled Text vs ChatGPT for handling structured, multi-part scene instructions specifically.
Where This Fits in Your Larger Process
None of this replaces good revision judgment — it just gives you a sharper tool for a specific, recurring problem. When you're doing your read-through pass and something feels hedged or oddly neutral in a scene that should have teeth, don't just reach for stronger adjectives. Check whether you handed the model a negative constraint without a replacement, and whether that constraint arrived in the same breath as the positive instruction it was quietly undoing. This matters at every stage, but it matters most in scenes carrying real weight — a confession under interrogation, a magic ritual that must not break its own rules, a first meeting between rivals who'll become something else by the end of the book. If you're troubleshooting an opening hook that isn't landing on page one, check whether you've told the model what your protagonist isn't doing without also telling it what fills that space — flat openings are often a symptom of exactly this cancellation problem. And when you send chapters out for outside eyes, it's worth flagging these restraint-heavy scenes specifically in your beta reader workflow for AI-assisted manuscripts, since readers often sense the flatness before they can name its cause. The habit worth building here is small but durable: every time you write "don't," ask yourself what replaces the thing you just removed. If you don't have an answer, the model won't either — and it'll pick the blandest possible option by default. Give it somewhere specific to put the tension instead, and the scene stops hedging and starts committing.