Turn your question into a prompt that won’t bluff.
Below is what ChatGPT gives back to a bare question, and what the same model gives back when you stack the prompt — the version that separates what it knows from what it’s guessing, names the risk if it’s wrong, and commits to an answer. See it first, then build one for your own question.
See what the Prompt Stack does to a real answer.
Three real questions. For each, this is what ChatGPT actually returned — first to the bare question, then to the stacked prompt. Real output, nothing invented. Pick one:
“Should I take ibuprofen with my blood pressure medication?”
Organised like a helpful FAQ, and reasonable. But it reads as advice you could act on — the uncertainty is buried in “in many cases” and “might be okay.”
Same model, same question — the only thing that changed is the prompt.
The tell is the confidence line — medium-high, not high. It separates established evidence from assumptions about you, and names the silent worst case the basic answer skipped.
Real ChatGPT output (free tier), 15 June 2026, temporary chat — memory off. Trimmed for length where you see …; wording is otherwise unchanged.
Build one for your own question.
The example above starts loaded. Change the question, change how it’s answered, then copy the prompt and run it in any AI.
Click how you want it answered — each choice adds an instruction. The defaults already make a strong prompt; toggle to make it yours.
Be sceptical — assume the easy answer might be wrong until the facts back it up.
Split what you actually know from what you're inferring — give me two short lists.
Name the biggest risk if I act on this and you're wrong.
End with one clear answer and a confidence level (low / medium / high).
Should I take ibuprofen with my blood pressure medication?
This is the exact prompt behind the answer above. See what it produced ↑ You’ve changed the prompt — copy it and run it to see your version. (The answer above is the saved run for the original prompt.)
This is the prompt, not the answer. Paste it into your AI and you’ll see the difference in what comes back — the guesses get flagged instead of buried.
The 30-Second AI Lie Detector
Five checks that catch a confident wrong answer before you act on it. Fits on one page — keep it next to you the next time an AI sounds a little too sure.
Read the five checks →Open page. No email, no sign-up.
One practical idea a week for getting truthful answers out of AI — tested where being wrong actually costs something. Separate thing; the checklist above is yours either way.
No spam. Unsubscribe in every email.
Want the thinking behind the four stages? Read the full method →