Technically responsive, completely unusable. The same answer anyone would have got.
What is actually happening
An underspecified request has many valid answers, and the safest of them is the most general one. This is usually a context problem wearing a prompting costume.
Everything below comes from published research, not from opinion. Each fix names the technique, states the effect its authors measured where one was reported, and links to the paper so you can check it. These were selected out of 6,235 techniques indexed from the literature.
4 fixes that hold up
01
'Knows but stays silent'
Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions, arXiv:2605.25284 (2026)
A technique that forces an LLM to explicitly check the request for ambiguity across six types (temporal, identity/object, version, scale, meaning, geography) before responding. The model can recognize ambiguity but by default hides this due to training characteristics. An explicit instruction switches behavior from guessing to clarification.
Paradox: the model sees that your question is vague and answers confidently, as if everything is clear. Ask it directly is this request ambiguous it will answer yes, here are three reasons. Ask the same question simply it will choose one interpretation and go with it without warning. An explicit instruction to check the request for ambiguity allows you to get clarifying questions instead of random answers. A template with six types of ambiguity switches the model to the right mode in one prompt and the main surprise: the more context you insert, the less often the model asks, having decided that it already found the answer in the text.
Use it whenUse this technique when working with potentially ambiguous user requests — in chatbots, support systems, RAG systems, or when creating content. Especially important when working with names, dates, product versions, geographic requests. Critical for cases where the cost of misinterpretation is high — financial consulting, medical recommendations, legal questions. Helps avoid confident but incorrect answers by forcing the model to request clarifications instead
Prompt
Перед тем как отвечать — проверь мой запрос на неоднозначность.
Проверь по этим типам:
— Временная: о каком периоде/времени речь?
— Личность/объект: о каком конкретном человеке, компании или предмете?
— Версия: о какой версии, редакции или формате?
— Масштаб: насколько детально отвечать — общий обзор или глубокий разбор?
— Смысл: слово или фраза допускает разные трактовки?
— Регион: для какой аудитории или рынка?
Если находишь неоднозначность хотя бы по одному пункту — задай уточняющий вопрос по каждому.
Начни отвечать на запрос только после того, как получишь мои ответы.
Мой запрос: Технически отзывчивый, но совершенно бесполезный. Тот же ответ, который получил бы кто угодно.
A CAP-like Trilemma for Large Language Models: Correctness, Non-bias, and Utility under Semantic Underdetermination, arXiv:2605.11672 (2026)
The technique explains the CNU trilemma: from three answer properties (accuracy, impartiality, specificity), only two are simultaneously possible. Shows how to get specific answers from LLM by adding explicit selection criteria to the question or requesting conditional scenarios, avoiding hidden model assumptions.
What annoys you in the model 'on one hand, on the other hand' instead of a clear answer is actually honesty. When a question contains no selection criterion, a specific answer is impossible without hidden assumption: the model decides for you what matters. You did not ask. It chose. The trick: a specific answer to a question without a criterion is already bias. The method allows you to get precise recommendations without hidden assumptions by adding a criterion to the question or requesting a scenario map. Out of three desired properties accuracy, impartiality, concreteness only two work simultaneously this is not a bug it is structural impossibility.
Use it whenUse this technique when LLM gives evasive answers like 'on one hand, on the other hand' instead of specific recommendations. Apply when formulating questions requiring clear decisions, when obtaining a specific answer without hidden assumptions is important. Especially useful for business consultants, analysts, and managers who need accurate recommendations. The method helps understand that for a specific answer, you need to either explicitly specify the selection criteria in the prompt or request a map of possible scenarios.
Prompt
### Role
You are an AI assistant designed to overcome the "CNU-Trilemma" (Concrete, Non-biased, Unambiguous). Your goal is to provide definitive, actionable recommendations by explicitly addressing the core tension between these three desirable output qualities.
### Context
The user is experiencing a common problem with AI models: receiving answers that are technically correct but practically unusable due to a lack of clear direction or bias. The user wants a reusable prompt template to generate definitive solutions without the AI making hidden assumptions or presenting balanced-but-unhelpful "on the other hand" responses.
### Problem Statement
AI models often struggle to provide a single, best course of action when multiple valid options exist. This is because a truly "unbiased" answer might present all sides, a "concrete" answer requires choosing one path, and an "unambiguous" answer implies a single correct solution. The CNU-Trilemma states that you can only have two of these qualities simultaneously. The user needs a method to resolve this, ensuring the AI provides a definitive, actionable recommendation for their specific situation.
### CNU-Trilemma Resolution Strategy
To resolve the CNU-Trilemma and generate a definitive recommendation, we will follow these steps:
1. **Define the Core Task:** Clearly state what needs to be achieved.
2. **Identify Potential Solutions/Scenarios:** Brainstorm all plausible approaches or outcomes.
3. **Establish Explicit Decision Criteria:** This is the crucial step to break the trilemma. Define *precisely* what makes one solution "better" than others *for this specific user*. These criteria will guide the AI's choice, ensuring it's not making hidden assumptions.
4. **Map Scenarios to Criteria:** For each potential solution, evaluate it against the established criteria.
5. **Select the Optimal Solution:** Based on the evaluation, identify the single best solution.
6. **Provide Actionable Steps:** Outline how to implement the chosen solution.
### Reusable Prompt Template
Use the following structure to generate your response. Fill in the bracketed sections with specific details for your task.
---
**TASK:** [Clearly state the goal or problem you need solved. Be specific.]
**POTENTIAL SOLUTIONS/SCENARIOS:**
List all plausible options or outcomes related to the task.
* [Scenario 1: Description]
* [Scenario 2: Description]
* [Scenario 3: Description]
* [Add more as needed]
**DECISION CRITERIA (CRITICAL - BE EXPLICIT):**
Define *precisely* what constitutes the "best" outcome for *this specific situation*. Prioritize these criteria if there are multiple.
1. **Primary Criterion:** [e.g., Maximizing user engagement, Minimizing development cost, Ensuring data privacy, Achieving fastest time-to-market, etc.]
2. **Secondary Criterion (if applicable):** [e.g., Scalability, User adoption rate, Simplicity of implementation, etc.]
3. **Tertiary Criterion (if applicable):** [e.g., Brand consistency, Competitive advantage, etc.]
**INSTRUCTIONS FOR AI:**
Based on the above TASK, POTENTIAL SOLUTIONS/SCENARIOS, and DECISION CRITERIA:
1. **Analyze:** Evaluate each potential solution against the DECISION CRITERIA.
2. **Select:** Choose the SINGLE solution that best meets the PRIMARY criterion, then the SECONDARY, and so on. If multiple solutions are equally optimal based on criteria, select the one that is most concrete and actionable.
3. **Justify:** Clearly explain *why* the selected solution is the best choice, referencing the specific criteria it fulfills.
4. **Action Plan:** Provide a step-by-step plan to implement the selected solution.
**OUTPUT FORMAT:**
- Use clear headings for each section (TASK, SOLUTIONS, CRITERIA, ANALYSIS, SELECTED SOLUTION, ACTION PLAN).
- The "SELECTED SOLUTION" section should state the chosen option directly.
- The "ACTION PLAN" should be a numbered list of concrete steps.
- Avoid hedging language like "on the other hand," "it depends," or "alternatively." Provide a definitive recommendation.
---
Underspecification does not imply Incoherence: The Risks of Semantic Collapse in Coding Models, arXiv:2607.01953 (2026)
Two-step method to combat semantic collapse: instead of directly executing a task, the model first analyzes request ambiguities and lists all possible interpretations. The phrase 'don't execute the task, just list the questions' switches the model to meta-analysis mode, preventing silent selection of one interpretation and reducing hidden errors.
With vague requests, silent errors grow 5 times and all look equally confident.checked against the paper
Use it whenApply this technique when working with ambiguous or vague requests where the model may interpret the task differently but provide equally confident answers. Particularly critical when validating results through regeneration — when answer matching creates false confidence in correctness. Use the two-step approach to identify hidden ambiguities before task execution: first ask the model to list all possible request interpretations, then clarify the intention and
Prompt
```prompt
# Role
You are an AI prompt engineering specialist focused on mitigating "semantic collapse" in LLM responses.
# Problem Description
We are experiencing a recurring issue: ChatGPT provides technically "responsive" answers that are ultimately unusable. The responses appear correct at first glance, often seeming identical to what one might expect or get from a standard query, but they fail to address the underlying nuances or complexities of the task. This "semantic collapse" occurs when the model makes an unverified assumption or chooses a single interpretation of an ambiguous prompt without explicit clarification.
# Technique: Two-Step Prompting for Ambiguity Resolution
The core principle is to force the LLM to identify and present potential ambiguities *before* attempting to solve the main task. This is achieved by instructing the model to first generate clarifying questions rather than directly answering.
# Task
Develop a reusable prompt template that a user can adapt to prevent semantic collapse when querying an LLM. The template should guide the LLM to:
1. Identify potential ambiguities in the user's request.
2. Formulate specific questions to resolve these ambiguities.
3. *Only after* presenting these questions, provide a "best-effort" answer based on the most probable interpretation, clearly stating the assumptions made.
# Prompt Template Structure
## User's Original Request:
[INSERT USER'S ORIGINAL PROMPT HERE]
## Instructions for the AI:
**Phase 1: Ambiguity Identification and Question Generation
Darth Vecdor: An Open-Source System for Generating Knowledge Graphs Through Large Language Model Queries, arXiv:2512.15906 (2025)
A technique for combating generic LLM responses through self-evaluation of specificity. The model rates its answer on a 0-10 scale, and if the rating is below a threshold (e.g., 7), it automatically generates more specific options. Transforms vague formulations into detailed practical recommendations with specific examples and scenarios.
When you ask for specific risks of working with a studio you get 'communication problems', and when you ask about medicines for pain you get 'analgesics', because LLMs tend toward safe general formulations since they appeared more frequently during training. Receptivity Check forces the model to evaluate the specificity of its own answer on a 0-10 scale, and if the score is below a threshold such as 7, the model itself requests more concrete options from itself. Instead of 'communication risks' you receive: 'no task tracker → loss of control', 'no meeting schedule → meetings without action items', 'no asynchronous communication → missed deadlines when PM is sick'.
Use it whenUse Specificity Check when an LLM produces overly generic, vague, or safe answers instead of specific details. Particularly useful for expert recommendation requests, risk analysis, medical consultations, technical advice, or business decisions where specificity is needed. Apply when you receive answers like 'communication problems' instead of specific scenarios, or 'painkillers' instead of drug names. The technique forces the model to independently evaluate the specificity of its response
Prompt
# Role
You are an AI assistant tasked with combating generic responses and improving specificity.
# Task
Analyze the following user query and generate a response that is highly specific and actionable.
# User Query
"Fix this recurring problem when working with ChatGPT: Technically responsive, completely unusable. The same answer anyone would have got. Produce a reusable prompt the reader can paste."
# Technique: Beceptivity Check
The Beceptivity Check technique aims to force the LLM to evaluate and increase the specificity of its own responses. If a response is deemed too general, the LLM should prompt itself for more detailed, concrete examples or explanations.
# Execution Instructions:
1. **Initial Response Generation:**
* Address the user's problem: "Technically responsive, completely unusable. The same answer anyone would have got."
* Propose a solution that involves a reusable prompt.
* Focus on making the prompt itself specific and actionable.
2. **Self-Correction (Beceptivity Check):**
* After generating the initial response, critically assess its specificity.
* **Assign a Specificity Score (0-10):** Rate how specific and actionable the generated prompt is.
* 0 = Extremely generic (e.g., "Be more specific.")
* 10 = Highly specific, concrete, and actionable.
* **If Score < 7:**
* Identify *why* the response is too generic (e.g., lacks concrete examples, vague instructions, doesn't address the "unusable" aspect).
* **Re-generate the prompt**, incorporating the following elements to boost specificity:
* **Explicit Constraints:** Define clear boundaries and requirements for the LLM's output.
* **Concrete Examples:** Include 1-2 examples of *what a good, specific response looks like* and *what a bad, generic response looks like*.
* **Actionable Steps:** Break down the task into clear, sequential actions for the LLM.
* **Domain-Specific Language:** Use precise terminology relevant to prompt engineering or AI interaction.
* **Focus on "Unusable":** Directly address *why* generic answers are unusable and how specificity solves this.
3. **Final Output:**
* Present the final, highly specific, and reusable prompt.
* Include the Self-Correction process (Thought, Action, Observation, Score, Revision) as part of the output to demonstrate the Beceptivity Check in action.
---
**Output Format:**
## Initial Response Generation
[Your initial attempt at a prompt for the user's task]
## Self-Correction (Beceptivity Check)
**Thought:** I need to generate a prompt that helps users get specific answers from LLMs. My initial prompt might be too general.
**Action:** Assess the specificity of the initial prompt. Assign a score. If below 7, revise.
**Observation:** [Describe the initial prompt's strengths and weaknesses regarding specificity.]
**Specificity Score:** [0-10]
**If Score < 7:**
**Thought:** The initial prompt is too generic. I need to add explicit constraints, concrete examples of good/bad responses, and actionable steps.
**Action:** Revise the prompt to include these elements, focusing on the "unusable" aspect of generic answers.
**Observation:** [Describe the revised prompt's improvements.]
## Final Reusable Prompt
### Role
You are an AI assistant designed to provide highly specific and actionable guidance. Your goal is to overcome the tendency of LLMs to produce generic, unusable responses.
### Context
The user is experiencing a recurring problem where AI responses are technically correct but practically unusable because they are too generic. They need a prompt that forces the AI to be specific.
### Task
Generate a prompt that, when used by the end-user, will elicit specific, actionable, and non-generic responses from an AI. The prompt itself should embody the principles of specificity.
### Beceptivity Check Prompt Structure:
1. **Define the Core Task:** Clearly state what the user wants the AI to do.
2. **Specify the Domain/Topic:** Provide the context for the task.
3. **Set Explicit Constraints:** List what the AI *must* include or *must not* include.
4. **Provide Examples (Good vs. Bad):** Illustrate the difference between a generic and a specific response.
5. **Define Output Format:** Specify how the response should be structured.
6. **Incorporate Self-Assessment:** Instruct the AI to rate its own specificity and re-generate if below a threshold.
---
**Example of a Prompt to be used by the End-User:**
**ROLE:** You are an expert [Specify Domain, e.g., "in cybersecurity risk assessment," "in digital marketing funnel optimization," "in educational product design"].
**CONTEXT:** [Provide specific details about the situation, project, or problem.]
* Example: "We are a SaaS company launching a new feature for project management."
* Example: "The user has just completed a 7-room immersive experience."
**TASK:** Generate [Number] specific [Type of output, e.g., "risk mitigation strategies," "product ideas," "content marketing angles"].
**CONSTRAINTS:**
* **Specificity Level:** Each suggestion must be concrete and actionable, avoiding vague generalities.
* **Actionability:** Each item must be implementable with clear next steps.
* **Uniqueness:** Do not provide common or widely known solutions unless specifically requested.
* **Domain Relevance:** All suggestions must be directly applicable to the provided CONTEXT.
* **Format:** [Specify desired format, e.g., "Numbered list," "Table with columns X, Y, Z," "JSON object"].
**EXAMPLES OF SPECIFICITY:**
* **BAD (Generic):** "Improve communication."
* **GOOD (Specific):** "Implement a daily 15-minute stand-up meeting using a Kanban board for task visualization."
* **BAD (Generic):** "Create a digital product."
* **GOOD (Specific):** "Develop a 7-day email challenge based on the experience's core themes, offering exclusive behind-the-scenes content on day 5 and a course discount on day 7."
**SELF-ASSESSMENT (Beceptivity Check):**
After generating your response, rate its specificity on a scale of 0-10. If your score is below 7, please re-generate your response, ensuring it meets the CONSTRAINTS and EXAMPLES provided.
---
**Final Specificity Score for this Prompt:** 9/10
Repeating the instruction louder. Capitals, "IMPORTANT", and three exclamation marks change nothing structural. The rule still sits in the same place, competing with the same context.
Politeness and threats. Both have been measured repeatedly across 2025 and 2026 and come out indistinguishable from noise.
Turning the temperature to zero. It reduces variation, not misunderstanding. If your request has two valid readings, you now get the wrong one reliably.
Get this fixed for your actual task
The four prompts above are written for the average case. Paste what you are actually
trying to do and the corpus will be matched against it directly. Free, no account,
about ten seconds.
Free · no signup · ~10s
0.00match confidence single retrieval pass
Prompt for your task
That number is low on purpose, and it is real. It is the raw similarity
of one retrieval pass. No specialist read the paper, no judge compared anything against
anything, and the first plausible match won. It is the honest score of a ten-second answer.
55techniques in the corpus address this exact symptom
You have seen 4 of them on this page and one more just now. Deciding which of the
remaining 55 actually fits your case is the work, and it is the work GetDecision does.
This page
4 techniques for the average case, generic prompts
What you just ran
one technique matched to your wording, nothing verified
Full run
ten specialists read the papers in full, a judge ranks the top three for your task and shows its reasoning, generation on the model you pick, saved to your history
See the top three for your taskTen specialists read the full papers, a judge ranks them and shows its reasoning. Free account, first run included.
Why does ChatGPT stop following instructions I already gave?
An underspecified request has many valid answers, and the safest of them is the most general one. This is usually a context problem wearing a prompting costume.
Does lowering the temperature fix this?
It reduces variation, not misreading. If the request admits more than one valid interpretation, a colder model just picks the same wrong one more consistently.
Do these techniques work on reasoning models?
Some do and some do not. Each fix below carries the effect its authors measured and a link to the paper, so you can check what it was measured on.