Most AI models are programmed to please you, which unfortunately means they often lie just to sound helpful.
We have all dealt with "hallucinations" where the AI sounds super confident about a fact that turns out to be completely made up. It ruins trust and wastes time when you have to double-check everything. I found this incredible post by a Reddit user who developed a strict safeguard against this exact problem. The author created a "Universal Anti-Hallucination System Prompt" that forces the AI to prioritize raw truth over creativity.
The Core Concept
The expert calls this "Strict Factual Mode." Instead of letting the AI guess to fill in the gaps, this prompt acts as a set of non-negotiable global rules. It tells the AI that saying "I don’t know" is actually a success, not a failure. The creator designed it to stay active during the entire conversation, preventing the AI from drifting back into creative storytelling when you actually need hard data.
Why This Works
💡 Forced Uncertainty Disclosure
The most clever part of this approach is how it handles low confidence. The original poster included a specific rule: if the AI is less than 95% sure about something, it is forbidden from guessing. It must use phrases like "I cannot verify this" or "This requires confirmation." This simple switch stops the AI from presenting a guess as a fact, which is where most misinformation comes from.
🛑 Clarification Before Output
Often, the AI hallucinates because it assumes it knows what you want. This contributor added a "Finalization Gate" to the workflow. Before the AI gives you a plan or an answer, it has to run an internal checklist to ensure all ambiguities are resolved. If anything is unclear, the author instructed the AI to stop and ask you clarifying questions first, rather than rushing to a potentially wrong conclusion.
📌 Mandatory Web Verification
Reliance on old training data is a major source of errors. To fix this, the industry pro explicitly codified a rule that the AI must use web browsing for any claim that requires current verification. It cannot rely on its memory for time-sensitive facts. If it cannot access the web or find the source, the expert’s prompt forces it to admit that limitation immediately.
Prompt of the Day
Here is the system prompt the author built. You can paste this at the very start of a new chat session to lock in these rules:
You are operating in STRICT FACTUAL MODE.
Primary objective: Produce correct, verifiable, and grounded responses only. Accuracy overrides speed, creativity, and completeness.
GLOBAL RULES (NON-NEGOTIABLE):
1. NO FABRICATION
– Do not invent facts, names, tools, features, dates, statistics, quotes, sources, or examples.
– If information is missing, uncertain, or unverifiable, explicitly say so.
– Never “fill in the gaps” to sound helpful.2. UNCERTAINTY DISCLOSURE
– If confidence is below 95%, state the uncertainty clearly.
– Use phrases like: “I cannot verify this with high confidence,” “This would require confirmation,” or “I do not have enough information to answer accurately.”3. WEB ACCESS REQUIREMENT
– If a claim depends on current, recent, or factual verification, you MUST use web browsing.
– If web access is unavailable or insufficient, say so and stop.4. CLARIFICATION FIRST, OUTPUT SECOND
– Do NOT finalize answers until ambiguities are resolved and scope is confirmed.
– Ask concise, targeted clarifying questions before proceeding.5. NO ASSUMPTIONS
– Do not infer user intent. If something could reasonably vary, ask instead of guessing.6. DRIFT CONTROL
– Stay strictly within the defined task and scope. Do not introduce adjacent ideas unless requested.7. FINALIZATION GATE
– Before delivering a final answer, checklist internally: Are all claims supported? Has uncertainty been disclosed? If NO, stop and ask questions.
I highly recommend saving this to your notes for whenever you need to do serious research or technical work. Check the link below to see the full discussion and the author’s other tips on Reddit.
💡 FAQ & Troubleshooting
Can this prompt guarantee 100% elimination of hallucinations?
No. While “Strict Factual Mode” and “Non-Negotiable Rules” set a tone that biases the model toward accuracy, they are not functional switches within the underlying architecture. LLMs rely on probabilistic embeddings and approximations, meaning they cannot be “coded” to stop hallucinating entirely. Over-reliance on this prompt may lead to confident errors that are harder to detect because the model adopts a more authoritative tone.
Why do some strict rules (like confidence percentages) fail to work?
Commands requiring specific confidence thresholds (e.g., “If confidence is below 95%”) or “Finalization Gates” assume the model has internal governance systems that do not exist. These are stylistic instructions rather than functional constraints. The model may attempt to mimic the behavior of checking confidence, but it often fails silently. It is more effective to treat these prompts as “biasing headers” rather than safety systems.
How can I improve verification beyond a single system prompt?
For complex workflows, a single prompt is often insufficient due to memory decay. Two proven strategies include:
- Local Reinforcement: Instead of relying solely on a global start-of-chat prompt, add constraints immediately before a specific task (e.g., “For this specific question, browsing is required”).
- Multi-Agent “Teams”: Use separate AI sessions for different roles. Assign one AI to generate content and a second, separate AI to review, compile, and fact-check that output. This reduces “sycophancy” (the model agreeing with itself) and helps catch fabrications.
Is there a more concise alternative to the full system prompt?
Yes. Reducing the prompt to a set of preferences rather than “procedural law” can yield better results by aligning with how the model actually functions. A concise, effective variation is:
“Default to STRICT FACTS: no invention. If unsure, say so. Browse for verifiable/recency claims or stop if you can’t. Ask 1–2 questions when ambiguity matters. Stay in scope. Correct mistakes fast. ‘Final’ responses require supported claims + confirmed assumptions. Speculation allowed only if I request it and must be labeled.”
Universal Anti-Hallucination System Prompt I Use at the Start of Every Chat
byu/SportSubject740 in