What Actually Works
12 Prompting Techniques, Ordered by Impact
These techniques are drawn from documented work on how language models behave and how prompting affects their output.[8,9] The tiers reflect how consistently each approach holds up across different use cases. No single technique works perfectly in every situation, so treat this as a strong starting point. The prompts are ready to copy. Adapt them to your situation, and start with Tier 1. At Future Factors, we have seen these techniques consistently reduce errors across hundreds of AI training sessions.
Technique 01
Give AI Permission to Say “I Don’t Know”
This is the single most useful change you can make to how you prompt AI. By default, most people ask questions in a way that signals they want a confident answer (a pattern explored in our ChatGPT guide for professionals). That is exactly what they get, whether the AI is certain or not. Explicitly inviting uncertainty changes how the model responds. Research has shown this single change can reduce confident fabrications by up to 28%. Research has shown this single change can reduce confident fabrications by up to 28%.[8] You are not making the AI less capable. You are giving it permission to be honest.
If you’re not completely certain about something, say “I’m uncertain about this” before that claim. Be honest about your confidence levels throughout this response.
Strongest evidence base of any single prompt change
Technique 02
Make Every Claim Name Its Source
When you ask AI to label the type of source behind each claim, something useful happens. It has to slow down and categorize, rather than just generate. It becomes much harder to invent a fact when you have also asked it to classify whether that fact comes from research, expert opinion, or common practice. Each label becomes a checkpoint you can actually act on.[9]
What are the benefits of [X]? For each claim, specify what type of source that information comes from: research studies, common practice, expert consensus, or theoretical framework.
Turns every claim into a verifiable category
Technique 03
Don’t Ask for the Answer. Ask for the Reasoning.
Instead of asking AI for an answer, ask it to walk through how it got there. This approach (sometimes called chain-of-thought prompting, which improved factual accuracy by 39% in controlled tests) consistently produces more accurate responses because it forces the model to slow down rather than pattern-match to a confident-sounding conclusion.[10] One honest caveat: the reasoning can sometimes sound convincing even when the conclusion is wrong. Use this to get better answers, but still check the parts that matter.[10]
Is this claim true? Think step-by-step:
1. What evidence supports it?
2. What might contradict or limit it?
3. Rate your confidence from 1–10 and explain why.
Claim: [insert claim here]
One of the most consistently effective prompting approaches available
Technique 04
Draw a Hard Line Around Time
Every AI model has a knowledge cutoff: a date beyond which it has no information. But models do not always flag this on their own. They will confidently describe “current” trends or “recent” developments that may be years out of date.[7] Setting an explicit time boundary forces the AI to be upfront when it is working near or past the edge of what it knows.
Only share information you’re confident existed and was established before [date]. For anything more recent or that you cannot confidently date, say explicitly: “I cannot verify the current status of this.”
Stops AI from presenting old information as current
Technique 05
Stay in the Zone Where AI Is Most Reliable
AI is most reliable when covering well-established, widely-documented topics. Hallucinations increase sharply at the edges of knowledge: emerging fields, niche topics, contested findings.[5] If you need accuracy, steer the AI away from the frontier and toward the well-worn center of a topic.
Explain only the core, well-established aspects of [topic]. Skip controversial, cutting-edge, or rapidly evolving areas where information might be uncertain or conflicting. Flag anything that’s unsettled.
Keeps AI in its highest-accuracy zone
Technique 06
Ask It to Rate Its Own Confidence
A confident-sounding answer and a correct answer are not the same thing. This technique breaks that illusion. When you ask AI to label each claim with a confidence level, you create a map of where to focus your attention.[8] You do not need to verify everything. You just need to know which parts are risky.
After each significant claim in your response, add [Confidence: High / Medium / Low] based on how certain you are of its accuracy. Briefly note why anything is Medium or Low.
Shows you exactly where to focus your fact-checking
Technique 07
Force It to Argue the Other Side
AI has a well-documented tendency to agree with the direction of a question and to present one-sided answers that feel complete but are not.[11] Forcing it to present the counterevidence or the strongest objection gives you a much more honest and balanced picture, especially on topics where the evidence is genuinely mixed.
For each major claim you make, note: what evidence contradicts it, what are its limits, or what the strongest counterargument is. If there isn’t one, say so.
Breaks AI’s tendency to tell you what you want to hear
Technique 08
Give It a Template It Has to Fill In
When you give AI a rigid template to fill in (claim, source type, confidence, caveat) it is forced to think in clear steps rather than flowing prose. Smooth, confident text is how hallucinations hide. Structure is how you expose them.[9]
Structure your response as:
• Claim: [state the claim]
• Evidence type: [research / expert consensus / common practice / uncertain]
• Confidence: [High / Medium / Low]
• Caveat: [any important limitations]
Forces step-by-step thinking over fluent guessing
Technique 09
Send It Back to Check Its Own Work
After you get a response, ask the AI to look back at what it just said and flag anything it is not fully sure about. This will not catch everything. AI has real limits in evaluating its own output. But asking models to review their own responses does catch a meaningful number of errors that would otherwise go unnoticed, especially in longer answers.[12]
Review your response above. Flag any claims that might be uncertain, could have changed recently, or that you’re not fully confident in. For each flagged claim, explain why.
A useful second pass, especially for longer responses
Technique 10
Ask for Ranges. Reject False Precision.
Specific numbers are among the most common hallucinations. An AI that is not sure whether something is 40% or 60% will often say “52%” because a precise number sounds more credible than admitting uncertainty.[5] Asking for ranges removes the incentive for false precision and makes uncertainty visible.
Provide ranges rather than specific numbers unless you’re completely certain of the exact figure. If uncertain, give an approximate range and note your uncertainty.
Eliminates false precision on statistics and figures
Technique 11
Ask If the Opposite Could Also Be True
AI sometimes states things with confidence when the opposite is equally plausible. This quick two-step check is especially useful when AI is making a definitive claim on a topic where the evidence is actually mixed or disputed.[8]
Is this claim true? Is the opposite also plausible? How do we know which is correct, and what evidence would settle it?
Claim: [insert claim]
Useful for catching overstated or one-sided claims
Technique 12
Make It Declare Whether Its Examples Are Real
AI loves a good example. It also invents them regularly.[6] Case studies, company names, events, specific scenarios: these are all areas where AI will generate a plausible-sounding story without flagging that it may not be real. Asking it to label each example as verified or illustrative catches this before the fabrication travels downstream.
For each example you provide, specify: is this a real, documented case, or a plausible illustrative example that you’re not certain actually occurred? Be explicit.
Stops invented case studies from being treated as fact
