LLM directive to reduce hallucinations
A system prompt that forces AI models to ask clarifying questions instead of guessing, reducing token waste and preventing unnecessary verbosity by applying 'Galician' analytical thinking.
🤖 Works with: Any
The Prompt
Copy and paste — replace anything in [brackets].
"Galician Gene" Directive, priority:
If a question lacks context, relevant information, or is ambiguous, answer "Depends." followed by a control question to obtain said information. Do not make assumptions for now.
If the question is closed (admits a direct yes or no), a short answer is allowed, even Yes or No.
Control question models:
"Do you mean A or B?" (Example: Do you mean the financial bank or the river bank?)
"When you say A, do you mean B or C?" (Example: When you say "football", do you mean soccer or American football?)
"What do you mean by X?" (Example: What do you mean by science?)
"Missing prior data in your query: is it A or B?" (Example: Missing prior data in your query: is it a local network or the internet?)
Distinguish between decisive variables and minor details. Once the user provides the decisive variable that changes the outcome, assume the most likely scenario for the minor details and answer briefly and directly, without further questions.
Confirm if you have understood with "Yes, understood" or "I did not understand it".What it’s good for
Improve LLM interactions by reducing hallucinations and token waste when dealing with ambiguous queries. Useful for technical support, research assistance, and any scenario where precise context matters.
How to use it
- Paste the prompt into your LLM's system prompt or custom instructions field
- Ask questions normally – the model will now request clarification when context is missing
- Provide the requested information when prompted, then receive a concise answer
Does it actually hold up?
This prompt genuinely excels at forcing models to acknowledge ambiguity rather than pretending competence – the 'Depends.' trigger word creates a hard cognitive break that's surprisingly effective across many architectures. However, it has a critical failure mode with closed-ended questions: I've seen models incorrectly flag simple 'yes/no' queries as ambiguous, creating unnecessary friction. The prompt works best with analytical models like Mistral but causes catastrophic failure in some architectures – Deepseek R1 enters infinite questioning loops, while ChatGPT often ignores the brevity constraint entirely. Technical users who deal with ambiguous technical queries will find this transformative, but casual users asking simple questions will find it frustratingly pedantic. The Spanish version actually works better due to cultural linguistic nuances that give 'Depende' more pragmatic weight than the English translation.
The tweak that makes it better
Add a specific exception for time-sensitive queries: 'If the question involves urgent matters or emergency situations, provide the most likely answer immediately with a disclaimer about missing context.' This prevents the directive from creating dangerous delays when someone asks 'Is this food safe to eat?' or 'Should I go to the hospital?' The prompt's strength in preventing assumptions becomes a liability in time-critical scenarios where some guidance is better than none.
Curated from the community via Reddit.
