How to Use Negative Prompts to Purge AI Prose Tics

Every writer who's used AI long enough has had the moment. You're reading back through a chapter and something feels... off. Not wrong, exactly. Just weirdly familiar. You scroll up and there it is again — another character who "felt a wave of something" or another sentence that starts with "The air was thick with." You check the previous chapter. Same thing. The chapter before that. Same thing. The AI has latched onto a phrase and it's been quietly seeding it through your manuscript like a weed.

These are AI prose tics: recurring syntactic patterns, stock phrases, structural habits, and tonal defaults that AI models fall back on when they're not actively steered away from them. They're not random. They're systematic. Every model has them, and they're different enough between models that a chapter drafted in one session can read differently from a chapter drafted in another, even when you're trying to keep the voice consistent.

The good news is they're suppressible. Not by vague instructions like "write more naturally" — that does almost nothing — but by building a targeted library of negative prompts: specific, evidence-based instructions that name the exact tics you want eliminated. Here's how to do it systematically.

What AI Prose Tics Are and Why They Accumulate

AI models are pattern-completion engines trained on enormous amounts of text. When you ask them to write fiction, they're drawing on statistical likelihoods — what phrase tends to follow what context. This means they have comfort zones. Sentence structures they return to. Emotional descriptors they favor. Transitional constructions they lean on when moving between scenes.

The most common tics fall into a few categories. There are phantom sensation phrases — "a shiver ran down her spine," "his chest tightened," "she felt something cold settle in her stomach." There are atmospheric filler constructions — "the silence stretched between them," "the room felt smaller somehow," "the air carried the weight of." There are echo openings, where multiple consecutive paragraphs start with the same syntactic structure ("She turned. She saw. She felt."). And there are dialogue attribution tics, where every line of dialogue that needs emphasis gets "said quietly" or "said with something unreadable in his eyes."

They accumulate across a draft because you're not writing in one sitting. Each new session with the AI is, in some sense, a fresh context window. The model doesn't remember that it used "the weight of unspoken things" four times last Tuesday. It just knows that in the current scene, with the current emotional register you've set up, that phrase is statistically likely. So it uses it again.

Tics aren't stylistic choices the AI is making. They're defaults it falls into when you haven't given it something more specific to reach for.

This is why generic instructions don't work. Telling the AI "avoid clichés" is like telling someone to "avoid bad driving" — technically correct, operationally useless. You need to name the specific roads you don't want them taking.

How to Audit Your Draft for Tic Patterns Before You Prompt

Before you can suppress anything, you need evidence. Gut feeling isn't enough here — you think you know what the tics are, but you're probably only catching the most obvious ones. The subtler structural habits are harder to spot in context because your brain smooths them over when you're reading for meaning.

The fastest audit method is a phrase extraction pass. Copy 3,000-5,000 words of AI-assisted draft into a plain text editor and run a simple word frequency count. Free tools like WordCounter.net or even a quick Python script will do it. You're not looking at single words — you're looking for repeated two-to-four word phrases. When you see "the weight of" appearing six times and "her chest" appearing nine times, you've found your tics.

After the frequency pass, do a sentence-opening audit. Go through your draft and write down the first three words of every paragraph. Just the first three. When you look at that list, patterns jump out in a way they don't when you're reading normally. You'll see that eleven paragraphs start with "She [verb]-ed" or that you have a cluster of paragraphs opening with "[Character name] [emotion verb]-ed." Those are structural tics, and they're often harder to catch than phrase tics.

Finally, do an emotional beat audit. Read through your draft and underline every place where a character's internal emotional state is described. Then look at what vocabulary is doing that work. You'll often find the AI has a narrow emotional lexicon for your protagonist — maybe she feels "a hollow ache" or things "settle in her chest" or she experiences emotions as physical sensations in her throat about 80% of the time. This tells you what emotional vocabulary you need to actively expand or replace.

Keep a running document. List each tic, the approximate frequency, and one or two examples from your draft. This becomes the foundation of your negative prompt library.

Building a Personal Negative Prompt Library from Your Audit Results

A negative prompt library is exactly what it sounds like: a saved, organized collection of suppression instructions that you can paste into your prompts when needed. The key word is personal. This library should be built from your actual draft evidence, not from generic lists of AI clichés you found online.

Structure your library in tiers. Tier 1 is universal suppression — tics so prevalent in your draft that you want them blocked in almost every prompt, across all scenes and characters. Tier 2 is context-specific suppression — tics that only appear in certain scene types (action sequences, emotional confrontations, atmospheric description). Tier 3 is character-specific suppression — habits the AI has developed for a specific character's voice or POV.

Here's what a real Tier 1 suppression block might look like after an audit, built for a draft of a psychological thriller in close third-person:

Notice that this prompt names specific phrases, names specific structural patterns (paragraph openings), and gives a replacement direction ("through specific thought or perception"). That last part matters. Suppression without redirection can produce flat, over-explicit prose. You're not just closing doors — you're pointing at a window.

For dialogue scenes specifically, you often need a different suppression block. Here's one built for a draft that had heavy dialogue attribution tics:

This one works because it's surgical. It's not saying "write better dialogue." It's saying: here are the three crutches you keep reaching for, don't reach for them, and here's what to use instead.

Build out your library over time. When you catch a new tic, add it. When a suppression instruction produces something flat or weird, revise the instruction. Treat it like a living document.

Stacking Negative Prompts Without Strangling Output Variety

There's a real failure mode here that you'll hit if you're not careful. You build a comprehensive suppression list, you paste in twenty constraints, and the AI produces something that technically avoids all the tics — but reads like it was written by someone under extreme stress. Over-constrained, jerky, slightly inhuman in a different way than before.

This happens when the suppression load gets too high relative to the creative space you've left open. The model is spending so much processing on what not to do that it has less room to do anything interesting.

Suppression is only as good as the alternative directions you give alongside it. Constraints work when they redirect, not just when they block.

The practical rule: for every two or three suppression instructions, include one generative direction. This isn't about balance for its own sake — it's about keeping the model anchored to something affirmative it can reach toward.

Also, be specific about scope. Don't paste your entire Tier 1 library into every single prompt. If you're writing a chase sequence, your emotional interiority suppression block is probably irrelevant — suppress what's likely to appear in this type of scene, not everything that's ever appeared in your draft.

Here's an example of a well-stacked prompt that has multiple suppressions without over-constraining:

Four suppressions, clear character direction, a specific structural instruction for the ending. This is manageable. The model has room to move.

You can also stagger your suppressions across sessions. If you have ten identified tics, don't suppress all ten in a single prompt. Suppress the three most prevalent ones consistently, then rotate in others based on what the current scene is most likely to trigger. Pay attention to which tics cluster in which scene types — your audit will have given you this information.

Testing and Calibrating Suppression Across Multiple AI Models

If you use more than one AI model — and most fiction writers do, at this point — you need to know something important: the tics are different between models, and so is the suppression response.

A suppression instruction that works perfectly in one model may produce strange results in another. Some models respond well to explicit lists of banned phrases. Others respond better to instructions framed as character or style directives. Some models need you to put suppression instructions at the end of the prompt; others perform better with them at the start.

Test your suppression library against your actual workflow. Pick a scene you're about to write. Write it with your full suppression block active. Then write it again with suppression removed. Read both versions. You're looking for two things: did the suppression actually eliminate the tics, and did it cost you anything in quality or variety?

If the suppressed version is visibly more mechanical, you've over-constrained. Cut back to your highest-priority suppressions. If the unsuppressed version has the tics but also has noticeably better moments, that's useful data — it tells you that specific suppression instruction is blunt and needs refinement.

For models that seem resistant to phrase-level suppression, try reframing. Instead of "do not use [phrase]," try building the suppression into a style description:

This reframes the suppression as a voice characteristic rather than a prohibition. For some models, this produces much better results because you're giving them a positive identity to write from rather than a list of negatives to navigate around.

Keep notes on what works in which model. A simple table — model name, suppression approach, result quality — will save you enormous time as your library grows. What you learn about one model won't transfer automatically, but the audit methodology and the library structure will.

The method here isn't complicated, but it does require the kind of systematic attention that's easy to skip when you're in the middle of drafting. The writers who end up with the cleanest AI-assisted prose aren't the ones who find the magic prompt — they're the ones who treat their drafts like evidence, read them like an editor, and build their suppression library from what they actually find rather than what they assume is there.

Start with your current draft. Run the phrase extraction pass tonight — it takes twenty minutes. Write down the five tics that show up most frequently. Build a Tier 1 suppression block from just those five. Use it in your next three writing sessions and watch what changes. That's the whole starting point. Everything else grows from that.