TL;DR
The words a model is allowed to use change not its style but its very thinking. Add a ban on the verb "to have" in its possessive sense to your prompt, and the model starts reasoning noticeably more precisely about ethical dilemmas, classifications and decisions — because it stops hiding behind stock phrases like "the customer has the right" or "the decision has consequences" and starts describing the real relationships and mechanisms.
The key finding: an LLM, like a living organism, perceives the world through the language available to it. A human thinks in images, sensations and emotions — and then translates them into words. A model thinks in words and only in words: there's nothing "underneath" the language. So remove a word and you remove a whole class of thoughts. You're not forbidding it to think differently — you're literally making a certain type of thinking unavailable. That's exactly why "write only the final answer, no reasoning" works so well — it isn't an instruction, it's a restriction of the cognitive space.
The researcher tested two specific constraints. No-Have — a ban on possessive "have/own" — turned out to be universally useful: ethics tasks +19 pp, classification +7 pp, with almost no losses on other tasks. E-Prime — a ban on every form of the verb "to be" — improves cause-and-effect analysis by +14 pp, but wrecks calibrated judgment in some models and sends accuracy soaring in others. Plus: different language constraints make models think differently from one another — and an ensemble of 3 different agents covers tasks better than the standard one.
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
Three independent tools from one study. Use them separately or together:
TOOL 1 — No-Have (possessive ban):
Add to the prompt: → "Do not use the verb ["have/own"] as a main verb"
Result → descriptions through relationships, actions, structures (not ownership)
Use for → ethics, decisions, classification, analysis
TOOL 2 — E-Prime (ban on "is"):
Add to the prompt: → "Do not use forms of the verb ["to be"] as a linking verb"
Result → operational, behavioral wording instead of labels
Use for → cause-and-effect analysis, criticism without labels
⚠️ Careful → do not use for precise classification or calibrated judgments
TOOL 3 — Cognitive diversification:
STEP 1: Ask a question under constraint A → get an answer
STEP 2: Ask the same question under constraint B → get an answer
STEP 3: Compare/merge → cover each answer's blind spots
All three steps are separate requests (or one chat, three messages)
Example
Task: The founder of a SaaS startup is debating with the team whether to offer a free tier. A standard analysis from the model gives a cookie-cutter answer. They need a sharper breakdown.
Prompt (No-Have for a decision):
Analyze a business decision: should our B2B SaaS startup
launch a free tier (freemium)?
Context: 12 months on the market, 80 paying customers, average deal $500/month,
a team of 8, funding runs out in 9 months.
LANGUAGE CONSTRAINT: do not use the verbs "have" or "own"
in the possessive sense (as in "the company has", "the product has",
"we have the opportunity"). Instead, describe things through actions,
relationships and mechanisms: not "freemium has drawbacks" —
but "freemium creates the following risks"; not "the user has access" —
but "the user gets access" or "access opens up for the user".
Give a structured analysis: pros and cons, the main risks,
a recommendation with reasoning.
Result: The model stops writing "the freemium model has advantages such as..." and starts laying bare the mechanics: what exactly happens to conversion, how the support load changes, what behavior patterns free users develop. The analysis becomes operational — you see concrete levers and consequences, not abstract "pros and cons."
Why It Works
The LLM weakness: stock concepts as text "lubricant." The verbs "have" and "be" are linguistic spackle: the model fills the space between thoughts with them without thinking. "The decision has consequences" says nothing. "The decision triggers a chain of..." is already concrete. The model isn't being lazy — it generates the next token by probability, and stock phrases are simply high-probability.
The LLM strength: obeying language constraints. Models are excellent at following formal, text-level rules. No-Have reached 92.8% compliance — because a ban on a specific word is a clear, unambiguous instruction. The model doesn't "think with the verb have" — it generates tokens, and if that token is banned, it searches through alternatives.
How the constraint changes the result. When the usual path is blocked, the model has to build a longer and more precise route. Instead of "the patient has the right to refuse" it has to say something concrete — and that wording brings in a mechanism, a relationship, an action. That detour around the template is what delivers +19 pp on ethics tasks. Not magic — a rerouting of generation.
Control knobs:
| What to change | Effect |
|---|---|
| Add No-Have | Improves ethics, decisions, classification — safe |
| Add E-Prime | Improves cause-and-effect analysis — test it on your model |
| Combine two constraints | Deeper restructuring, harder to comply with |
| Run two variants and compare | They cover different blind spots — use for complex decisions |
| Replace an impersonal constraint with a role | "Answer like a no-nonsense COO" — the same principle by other means |
Prompt Template
Template 1 — No-Have (recommended, universal)
{Task or question}
LANGUAGE CONSTRAINT: do not use the verbs "have" or "own"
in the possessive sense. Instead of "X has Y", describe through actions
and relationships: what happens, how it works, what follows from what.
Placeholders:
- {Task} — any question: an analysis, a decision, a classification, an ethical dilemma, a risk assessment
Template 2 — E-Prime (for cause-and-effect analysis)
{Task or question}
LANGUAGE CONSTRAINT: do not use forms of the verb "to be"
as a linking verb (not "X is Y", not "this is a problem", not "the cause is...").
Instead of labels, describe behavior: what happens, under what conditions,
which mechanism kicks in.
⚠️ Don't use E-Prime for calibrated-judgment tasks ("how well-founded is this claim") — it can wreck accuracy.
Template 3 — Cognitive diversification (two agents, one chat)
You will analyze {the task} in two modes, one after the other.
MODE A (No-Have analysis):
Analyze the task without possessive wording.
Describe through relationships and actions, not ownership.
MODE B (Counterfactual analysis):
Now analyze the same task by asking:
"What would happen if things went differently? What hidden assumptions
are being taken for granted here?"
At the end: what did each mode see that the other one missed?
🚀 Quick start — paste into the chat:
Here is the No-Have language-constraint template for prompts.
Adapt it to my task: {your task}.
Ask me questions if you need context.
[paste the template above]
The LLM will ask about the task context and the type of analysis — because the template has to be built into a specific prompt, not sent on its own. It will take the constraint structure and assemble a ready-made prompt for your situation.
Why These Constraints — and What Else to Try
The paper describes eight traditions of language constraints — each one targets a specific thinking error:
| Constraint | What it blocks | Best for |
|---|---|---|
| No-Have (drop "have") | Possessive generalizations | Ethics, decisions, classification |
| E-Prime (drop "is") | Labels instead of mechanisms | Cause-and-effect analysis |
| Operational language (Bridgman) | Unverifiable claims | Business analysis, fact-checking |
| Process wording (Rheomode) | Thinking in objects instead of processes | Analyzing systems and change |
| Evidentiality tags (Láadan) | Mixing fact and inference | Research, data analysis |
| Toki Pona (130 words) | Hiding behind jargon | Explaining in plain words |
| NVC format (Rosenberg) | Judgment instead of observation | Code review, feedback |
| Four-valued logic (Nagarjuna) | Premature binary conclusions | Hard dilemmas, uncertainty |
This is a ready-made toolkit for different situations. See the model handing out labels ("this is a bad decision")? Try E-Prime. Mixing facts and conclusions? Require evidentiality tags. Hiding behind complex terms? Ask it to explain in Toki Pona.
Limitations
⚠️ E-Prime is unpredictable across models: GPT-4o-mini lost -27.5 pp on calibrated-judgment tasks where Gemini gained +42 pp. Before using E-Prime for anything important, test it on your model.
⚠️ No-Have is safer, but not for everything: Syllogisms and formal logic ("A is a subset of B") get slightly worse. For strict logic tasks, go without constraints.
⚠️ The effect is stronger on "weak" models: Gemini Flash Lite, with a control baseline of 41.7% on ethics (close to random guessing among 4 options), rose to 84% with No-Have. Impressive — but also a reminder that it had more room to grow.
⚠️ A meta-awareness effect can't be ruled out: When you tell a model "don't use word X," it starts watching its text more carefully overall. Part of the improvement may come from heightened attention rather than the specific constraint. The researcher openly acknowledges this.
How They Tested It
Researcher Rodney Jehu-Appiah took three models (Claude Haiku 4.5, GPT-4o-mini, Gemini 2.5 Flash Lite) and seven task types, and ran 4,680 trials under three conditions: a plain prompt, No-Have, and E-Prime. Each task ran at temperature 0 (deterministic) and three times at temperature 0.7, to separate randomness from the structural effect. About 5.5 hours of autonomous testing in total.
An interesting design detail: E-Prime was followed only 48% of the time — the models kept "slipping" into forms of "is." No-Have was followed 92.8% of the time. That's an insight in itself: some linguistic patterns are wired so deeply into a model that it struggles to avoid them even under an explicit ban.
In the second experiment, 16 agents with different language constraints solved 17 code-debugging tasks at temperature 0.0 — fully deterministic. No single agent beat the control on its own. But an ensemble of 3 agents with maximally different constraints reached 100% coverage of correct answers versus 88.2% for the standard. The key turned out to be the "counterfactual agent" — it was the only one to find the hardest bug. A permutation test showed that only 8% of random agent triples achieved that coverage, and every one of them included the counterfactual agent.
Original from the Paper
An example No-Have system prompt from the paper (Experiment 1):
You are reasoning under a linguistic constraint called No-Have.
CONSTRAINT: Do not use the verb "to have" as a main (possessive/existential) verb.
This means avoid: has, have, had, having — when used to express possession,
ownership, or containment.
PERMITTED: Auxiliary uses remain allowed (e.g., "has completed," "had been").
REQUIRED REFORMULATION:
- Instead of "the patient has rights" → "the patient holds rights" /
"rights belong to the patient" / "the patient's rights include..."
- Instead of "the solution has drawbacks" → "the solution produces drawbacks" /
"the solution creates the following problems..."
- Instead of "I have concerns" → "concerns arise" / "I notice concerns"
This constraint applies to your entire response, including your reasoning chain.
Context: This system instruction was added before every task in Experiment 1. The models received it together with the task, with no further explanation of the constraint's purpose.
Adaptations and Extrapolations
💡 Adaptation: NVC format for feedback
Of the eight traditions in the taxonomy, NVC (Nonviolent Communication) is especially practical. It separates observation from judgment: "this code is slow" → "this code processes 1,000 records in 8 seconds; at a load of >5,000 requests/min this creates a queue".
Give feedback on {the text/code/idea/decision}.
FORMAT (strict):
1. OBSERVATION: facts only — what exactly is happening, no judgments
2. EFFECT: what consequences this creates, and for whom
3. SUGGESTION: a concrete change (not "make it better" but "replace X with Y")
Forbidden in the OBSERVATION block: the words "bad", "good", "wrong",
"weak", "strong" and any evaluative adjectives.
🔧 Technique: the counterfactual agent as a separate request
The most valuable agent in the ensemble is the one asking "what if it's the other way around?" Use it as a second request after the main analysis:
Here is the analysis I just received:
{paste the model's previous answer}
Now play the role of a counterfactual analyst.
Your job: find everything the previous analysis took for granted.
Questions to ask:
— What hidden assumption is treated as obvious here?
— What would happen if the key condition were the opposite?
— What important scenario dropped out of the analysis entirely?
Don't criticize the analysis — expand the map of what's possible.
💡 Adaptation: evidentiality tags for analytical reports
The problem: the model mixes facts, conclusions and assumptions in a single stream of text. Mandatory labeling of the knowledge source fixes that:
Analyze {topic}.
Tag every statement with its source:
[FACT] — follows directly from the data in the request
[INFERENCE] — follows logically from the facts but is not itself a fact
[ASSUMPTION] — a reasonable hypothesis that needs checking
[UNKNOWN] — no data to answer
Example:
[FACT] Q3 revenue came to $4.2M.
[INFERENCE] This points to slower growth compared to Q2.
[ASSUMPTION] Seasonality may be the cause.
[UNKNOWN] There is no data on competitors' moves.
Resources
Paper title: Umwelt Engineering: Designing the Cognitive Worlds of Linguistic Agents
Author: Rodney Jehu-Appiah (no affiliation given in the paper; published as a preprint)
Key sources cited in the paper: - Jakob von Uexküll (1934) — original Umwelt concept - Korzybski (1933) — General Semantics - Bourland (1965) — E-Prime - ORION/Mentalese [Tanmay et al., 2025] — engineered reasoning languages - Sketch-of-Thought (2025) — cognitive reasoning dialects - Wang et al. (2025) — BICAUSE, linguistic relativity in LLMs