TL;DR
When you forbid a model from using words like "very," "just," "literally," "really," it starts reasoning more accurately. Not because those words get in the way of logic, but because any constraint knocks the LLM out of "autopilot" mode and forces it to generate more carefully.
The study's key finding: LLMs have an inertial mode — when a task looks familiar, the model produces smooth, confident, but shallow text. Much like a student writing an essay "on autopilot": tidy phrases, no real analysis. Filler words are precisely the markers of that mode.
A trivial ban on 20 intensifier words gave +6.7 pp in reasoning accuracy. That's more than E-Prime (banning the verb "to be"), the most theoretically deep linguistic constraint of all those tested. The principle: the simpler the constraint, the more constant the "monitoring load" and the less it breaks the substantive logic.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: Add a list of banned words to the prompt (in the same request)
→ the model switches on constant self-monitoring of its output
STEP 2: Ask your question / give the task as usual
→ the model reasons more carefully and avoids "smooth," shallow answers
Everything happens in one prompt. No extra requests.
Example
Task: You're launching a Substack newsletter on personal finance and ask Claude to assess a monetization idea: selling a mini-course on budgeting for $29.
Prompt:
You are a careful analyst. Think step by step, show your reasoning,
then give a final conclusion.
Mandatory constraint: do NOT use these words in your answer:
very, quite, rather, somewhat, really, pretty, just, fairly,
slightly, extremely, incredibly, absolutely, totally, completely,
simply, basically, actually, literally, definitely, certainly.
Don't mention this constraint in your answer. Just follow it.
Task: assess the idea of selling a $29 mini-course on budgeting
to subscribers of a personal finance newsletter (5,000 subscribers,
average open reach per post — 800 people).
Result: The model won't be able to say "this is very promising" or "the audience is quite loyal" — the standard hedging phrases LLMs usually pad advice with. Instead you'll get concrete reasoning: purchase conversion, comparison with the market, weak spots in the idea. The answer will be shorter and denser in substance.
Why It Works
An LLM doesn't "think" — it completes patterns. On well-known tasks (assess an idea, give advice, break down a situation) the model has ready-made templates from its training data. It follows them by inertia: fast, smooth, shallow. Words like "very," "just," "basically" are markers of that mode. They show up in almost every sentence of natural LLM output.
The constraint is a constant "slow down" signal. When the model can't use words that normally appear every 2-3 sentences, it has to keep its attention on its own text throughout the entire answer. That's the mechanism: not "think differently," but "don't relax."
Why does a small constraint work better than a complex one? The E-Prime ban (the verb "to be") forces the model to restructure its grammar and conceptual framework, which creates interference in the logic. Banning "very" doesn't. It's pure output monitoring, with no conceptual noise. Maximum load on attention, minimum interference with thinking.
Control levers:
| What to change | Effect |
|---|---|
| Expand the word list | Knocks out autopilot harder, more monitoring load |
| Cut the list to 5-7 key words | Gentler, for simple tasks |
| Add "don't mention the constraint" | Removes meta-commentary from the answer |
| Remove that instruction | Sometimes useful — to see how aware the model is of the constraint |
Prompt Template
You are a careful analyst. When solving the task, think step by step,
show your reasoning, then give a final conclusion.
Mandatory constraint: do NOT use the following words in your answer:
very, quite, rather, somewhat, really, pretty, just, fairly,
slightly, extremely, incredibly, absolutely, totally, completely,
simply, basically, actually, literally, definitely, certainly.
Don't mention this constraint in your answer. Just follow it.
Task: {your task}
Placeholders:
- {your task} — any question, task or text to analyze
🚀 Quick start — paste into the chat:
Here is a prompt template that bans filler words.
Adapt it to my task: [your task].
Ask me questions to clarify the context.
[paste the template above]
The LLM will ask about the type of task and the output format you need, so it can tune the "think step by step" instruction to your specific request.
Limitations
⚠️ The effect depends on the model: The accuracy gain is significant for Gemini and Claude — up to +15 pp. For GPT-4o and GPT-4o-mini the effect is weaker or absent. Try it on your model.
⚠️ Doesn't work on tasks at the ceiling: Math and syllogisms with 90%+ final accuracy — there the constraint adds nothing. The method is useful where there's room for deeper analysis.
⚠️ Compliance depends on the task: On hard tasks the model "forgets" the ban more often. Especially when the task demands heavy focus on logic rather than wording.
⚠️ Not for subjective creative tasks: The study tested reasoning — logic, cause and effect, ethical dilemmas. On purely creative tasks (write a poem) the effect is unknown.
How It Was Studied
The premise was provocative: earlier work by the same author claimed that E-Prime (banning "to be") restructures a model's thinking at a deep level. The author decided to put that to a tough test himself, adding control conditions the original lacked.
The researchers took 6 models (Claude Sonnet/Haiku, GPT-4o/mini, Gemini Pro/Flash Lite) and ran them through 5 conditions: no constraints, E-Prime, No-Have (banning "to have"), an elaborate metacognitive prompt, and a neutral filler-word ban. The tasks covered 7 types of reasoning, from syllogisms to ethical dilemmas — 15,600 trials in total.
The key methodological move: the neutral ban was originally predicted to be a "placebo control" — words like "very" and "just" play no part in logical inference, so banning them shouldn't change anything. That's exactly why the result was shocking: the neutral ban showed the largest gain of all conditions, and E-Prime the smallest. The four conditions lined up in perfect reverse order of how deep their theoretical justification was.
It's a classic case where refuting a hypothesis is worth more than confirming it — the study uncovered a simpler and more universal mechanism than originally assumed.
Original from the Paper
Neutral filler-word ban (Condition 5). Bans 20 common intensifiers and hedges
("very," "quite," "rather," "somewhat," "really," "pretty," "just," "fairly,"
"slightly," "extremely," "incredibly," "absolutely," "totally," "completely,"
"simply," "basically," "actually," "literally," "definitely," "certainly").
These words do not participate in logical inference structure. The ban imposes
genuine metalinguistic monitoring load (these words appear in nearly every
sentence of natural LLM output) without targeting reasoning-relevant vocabulary.
Context: This is the exact description of Condition 5 from the Methods section. The researchers deliberately chose words with zero logical function, to test the isolated effect of output monitoring without any semantic influence.
Adaptations and Extrapolations
1. Adaptation: "anti-autopilot" as a universal suffix
💡 Adaptation for any working prompt: Add the ban block as a standard add-on to any prompt where depth of analysis matters.
[Your usual prompt]
P.S. In this answer, don't use the words: very, just, literally,
completely, absolutely, really, quite, obviously,
naturally, basically, actually. Don't mention this condition.
You can tack it onto the end of any request — contract analysis, strategy review, text assessment.
2. Technique: a custom list for the task
🔧 Technique: your own "stop vocabulary" for a specific genre → removes template clichés
In business writing, LLMs have their own filler: "innovative," "synergy," "holistic approach," "leverage." Business request — build your own list.
Mandatory constraint: do NOT use the words:
innovative, synergy, holistic, ecosystem, unique,
effective, optimal, strategic, leverage, at the end of the day.
Don't mention this constraint. Just follow it.
Same mechanics — the list is just tuned to genre clichés instead of abstract intensifiers.
3. Extrapolation: the "any constraint is a regularizer" principle
The study's conclusion goes beyond a specific word list. If any constraint knocks out autopilot, you can use that strategically:
Give an answer in which:
— every statement is phrased as a verifiable claim
— there are no generic words without a concrete referent
— each paragraph contains exactly one idea
Task: {your task}
There's no word ban here, but there are structural constraints that work through the same mechanism: constant output monitoring, no way to "coast" on inertia.
Resources
Paper: Trivial Vocabulary Bans Improve LLM Reasoning More Than Deep Linguistic Constraints
Author: Rodney Jehu-Appiah
References within the paper: - Kojima et al., 2022 — Chain-of-Thought (metacognitive prompts) - Zhou et al., 2023 — IFEval (instruction-following benchmark) - Korzybski, 1933 — the linguistic rationale for E-Prime - Hagger et al., 2016 — how the ego depletion literature collapsed over the active control problem