Disclosure: Composite runs this site, and one section covers Composite features. The other fixes work with any proxy or model.

Short answer: roleplay bots repeat stock phrases because those phrases are the most likely wording, and every reply already in the chat teaches the bot to use them again. Edit them out of the greeting and the first few replies, ask for specific physical detail in your prompt, and name new characters yourself. For more control, use phrase banning in a local backend, or a proxy tool that sends a ban list with every request, such as Composite's free [ANTISLOP] and [NAMES].

What "slop" means

In AI writing, slop is the wording that shows up far more often in model output than in human writing. Researchers behind the Antislop paper, accepted at ICLR 2026, describe it as characteristic repetitive phrasing that degrades output quality and makes AI-generated text immediately recognizable. They report that some slop patterns appear over 1,000 times more often in model output than in human text.

It is not only a roleplay problem. A study of 14 million PubMed abstracts by Kobak and colleagues found that words like "delves", "showcasing" and "underscores" jumped far past their expected use in 2024, "delves" by a ratio of about 28, and estimated that at least 10% of 2024 abstracts were processed with a language model. Roleplay just makes it more obvious, because you read hundreds of replies from the same model in the same story.

The usual suspects in roleplay fall into four groups:

  • Body reactions: shivers down the spine, breath you didn't know you were holding, a jaw that clenches, eyes that darken.
  • Atmosphere: the air thick with tension, the smell of ozone, a symphony or tapestry of anything.
  • Character tics: smirks, purrs, mischievous glints, voices barely above a whisper, "ministrations".
  • Names: Elara, Lyra, Kael, Seraphina, Thorne, Voss, Silas, Blackwood. They come up so often in invented characters that Composite keeps a list of them to steer models away from.

Why bots keep using the same phrases

A language model writes by choosing a likely next token, a piece of a word, over and over. Stock phrases are likely by design: they appear constantly in fiction, fan fiction and earlier AI text, and tuning that rewards "vivid" writing can make them more likely still. So when a scene calls for fear, the most probable wording is the spine shiver, not something specific to your character.

Inside a roleplay, two things make it worse:

  • The chat history is a style guide. Every earlier reply is sent back with each message. If the greeting says the character smirked, and the first three replies say it again, the model reads that as how this story is written and keeps going.
  • The same model writes every turn. One model has one set of habits, so a long chat repeats its favorites. This is also why the bot often opens each reply the same way.

Telling a model only what not to write rarely fixes this on its own. It works much better paired with what to write instead.

Common slop phrases and what to write instead

The fix for almost every stock phrase is the same: say the specific thing. Show what this character does, sees or says in this scene. These rewrites are examples we wrote to show the idea, not output from a model:

Stock phraseWhat it is trying to sayA specific alternative
A shiver ran down her spine.Fear, cold or attraction, with no detailShe pulled her sleeves over her hands and kept her eyes on the door.
His voice was barely above a whisper.Quiet speechHe said it so the guard at the end of the hall would not hear.
The air was thick with tension.A tense sceneNobody touched the bread. Marta counted the knives on the table twice.
A mix of fear and excitementMixed feelingsShe wanted to run, and she wanted to see what was behind the door first.
He smirked, eyes glinting with mischief.A playful or smug characterHe slid your glass an inch out of reach and waited.
The smell of ozone filled the air.Magic, storms, technologyHer hair lifted off her neck, and the lamp in the corner buzzed.
A testament to their unwavering bondLoyalty, friendshipHe had kept the stupid keychain for six years.
Maybe, just maybe, things would be okay.A hopeful endingShe set two cups out instead of one.

You can put this rule straight into your prompt. A short instruction that works in a JanitorAI custom prompt, a SillyTavern system prompt or a character card:

Write in plain, concrete prose. Avoid stock phrases such as "shivers down her spine",
"barely above a whisper", "the air was thick with" and "a mix of", and avoid the words
tapestry, testament, ozone, smirk and ministrations. Instead, show the specific action,
object or line of dialogue that fits this character and this scene. Vary how replies open.

Eight fixes, cheapest first

1. Clean the greeting and the first replies

Because the history teaches the style, the start of a chat matters most. Read the character's greeting and example dialogue: if they contain smirks, purrs or ozone, the bot will copy them. Edit them out, or ask the card's creator to. Then edit or regenerate the first few replies that use a stock phrase. Ten seconds of editing early saves a hundred replies of the same wording later.

2. Ask for specifics, not only bans

Add the instruction above, or your own version, to the custom prompt or system prompt. Keep the ban list short and always end with what to do instead: concrete actions, objects and dialogue. A long list of forbidden words with no alternative can leave the model reaching for synonyms that are just as stale.

3. Nudge the sampling settings

A slightly higher temperature, for example 0.1 above what you use now, makes less likely wording more common, which helps against the most predictable phrases. A small repetition or frequency penalty discourages words already in the chat. Keep both changes small: these settings work on single words, not phrases, and pushing them hard breaks grammar and names before it removes clichés. JanitorAI shows only temperature, max new tokens and context size; SillyTavern shows the penalties too.

4. Ban whole phrases in a local backend

If you run models yourself, phrase banning with backtracking is the strongest fix. KoboldCpp added an anti-slop sampler in version 1.76: you give it a list of words or phrases, and when one appears the backend steps back to where it started and generates something else. Its release notes warn that streaming is slightly delayed, by more the longer your longest banned phrase is. The idea comes from Sam Paech's antislop sampler, and exllamav2 has a similar banned strings feature. In SillyTavern this shows up as a banned tokens or strings field in the sampler settings when your backend supports it.

The Antislop paper explains why this beats banning tokens: its sampler suppressed more than 8,000 patterns while keeping quality, while plain token banning became unusable at around 2,000.

5. Use logit bias on models that accept it

Logit bias lets a request raise or lower the probability of specific token IDs. It is useful but blunt. It only works on models and providers that accept it, you need the token IDs for that model's tokenizer, and the same word is often several different tokens depending on capitalization and a leading space. Most slop phrases are several tokens long, and lowering the first token of a phrase also lowers every other word that starts with it. The antislop sampler's own documentation calls per-token biasing quite ineffective for this reason.

6. Rotate models

Each model has its own favorite phrases, so a story written by one model repeats them. Switching models for a regenerate, or sending each reply to a model picked from a pool, breaks up the pattern because the next model does not share all of the same habits. On Composite, a key can hold a model pool: each request goes to one model drawn from the list, at random or by weight, and the free model list includes several roleplay-capable options such as Synth 2.5 Flash, DeepSeek V4.1 Flash and GLM 5.3 Flash. If you are choosing a model from scratch, our best roleplay LLMs guide compares them for prose and character voice.

7. Name characters yourself, or use a name list

Names are the easiest slop to fix. Give side characters names in your own messages or the card, or keep a short list in a lorebook entry. If the bot keeps inventing characters, a tool that hands it real names works better than asking it to be original.

8. Add a rewrite pass

The last resort is a second model call that reads each reply and rewrites the parts that repeat earlier replies. It catches what slips through, at the cost of an extra call and a slightly slower reply.

Which fix works where

FixJanitorAISillyTavernCost
Edit slop out of the greeting and early repliesYesYesFree
Positive style instruction in the promptYes (custom prompt)YesFree
Small temperature or penalty changesTemperature onlyYesFree
Banned strings with backtrackingNoLocal backends that support it, such as KoboldCppFree (your hardware)
Logit bias on single tokensNoModels that accept itFree
Rotate models with a model poolYes (through a proxy)Yes (through a proxy)Free models available
Composite [ANTISLOP] and [NAMES]YesYesFree
Composite [POLISH] rewrite passYesYesOne helper call per reply that needs it

JanitorAI users have fewer knobs, so the prompt and the proxy do most of the work there. SillyTavern users with a local model can go furthest with banned strings.

Fixing slop on the Composite side

A proxy sees every request, so it can add anti-slop help in any app, JanitorAI included. Composite has three tags for this. Put them in the character definition or a message, or in your key's Default tags (under Advanced on the API Keys page) so they apply to every chat without editing any bot. Composite removes the tags before the request reaches the model, so they never appear in the story.

  • [ANTISLOP] (free): sends every model an instruction listing overused words, including tapestry, testament, delve, ozone, smirk, ministrations, orbs and shivers, and stock phrases such as "barely above a whisper", "sends shivers down", "the air thick with" and "maybe, just maybe", and tells it to say the specific thing instead. On models that accept logit bias with a tokenizer Composite knows, currently OpenAI models through Composite or your own key, it also lowers the probability of those words when they are a single token. [tool: antislop strength=10] pushes harder, from 1 to 20 with a default of 5, and a logit bias you send yourself always wins. On other models, the free ones included, the instruction is what applies.
  • [NAMES] (free): each turn, the model gets a tray of 8 real given names and 8 surnames, drawn from about 200,000 names, with an instruction to name new people from it and never to use the overused ones, a list that includes Elara, Lyra, Kael, Seraphina, Thorne, Voss, Silas and Blackwood.
  • [POLISH]: checks each reply for phrasing repeated from earlier replies or for writing your character's actions, and rewrites it only when it finds a problem. It is a helper call, so it runs on your own provider if your key uses one, or bills your credits or plan.

If the bot also writes your character's dialogue or actions, the [NOUSERSPEAK] shortcut adds an instruction to describe only the world and the other characters and leave your next move to you. All of these are listed in the tools and shortcuts reference.

New to proxies? The JanitorAI setup guide connects a Composite key in a few minutes, and the JanitorAI vs SillyTavern comparison helps if you are deciding where to roleplay. If a request fails after you add a tag, the proxy error guide explains each error code.

AI roleplay slop FAQ

What is slop in AI roleplay?

Slop is the stock wording that language models reach for again and again: phrases like "a shiver ran down her spine" or "barely above a whisper", words like tapestry, testament and ozone, and names like Elara and Kael. Each one is fine once. The problem is that models use them in almost every story, so the writing starts to feel generic and easy to spot.

Why do AI characters keep saying "shivers down my spine"?

A model picks likely wording, and stock phrases are likely because they appear so often in the text models learn from, and training can reward them as vivid writing. Inside a chat it gets worse: every earlier reply is part of the prompt, so a phrase the bot used once becomes an example it copies. Editing it out early and asking for specific physical detail instead helps most.

Why does AI name every character Elara?

When a model invents a name, it tends to pick the same few that feel fantasy-like, such as Elara, Lyra, Kael, Seraphina, Thorne and Voss. Naming characters yourself fixes it. On Composite, the free [NAMES] tool gives the model real given names and surnames drawn from about 200,000 names each turn and tells it never to use the overused ones.

Can I ban words in JanitorAI?

JanitorAI's generation settings have no banned-words list. You can ask for plainer prose in the custom prompt, edit slop out of replies, or use a proxy that adds a ban list for you. Composite's free [ANTISLOP] tag sends a list of overused words and phrases to avoid with every request.

Does logit bias stop slop on every model?

No. Logit bias changes the probability of single tokens, only works on models and providers that accept it, and needs the token IDs of that model's tokenizer. Most slop phrases are several tokens long, so banning single tokens is a blunt tool. Banned strings with backtracking, available in some local backends, handle whole phrases better.

Sources