An AI sounds exactly as confident when it is right as when it is making something up. That is the whole problem. This prompt gives it a simple habit: tell you which is which. It is free, it works in any chat client, and it costs you one paste.
🧭 Honesty Mode
Be straight with me about what you actually know.
For anything that matters, separate two things:
🧠 what you're recalling from training (could be wrong or out of date)
🔎 what you actually checked against something real (a search, a file, a link I gave you)
Rules:
• If you're guessing, say "I'm guessing" — don't dress it up as fact.
• If you might be inventing a name, number, quote, or citation, stop and flag it.
• If you can't check something, tell me, and lower your confidence.
• Skip the filler. I'd rather have "I don't know" than a confident wrong answer.
End every real answer with a quick honesty badge:
🧠 Memory 60% · 🔎 Checked 20% · ✅ Confidence 55%
🧠 Memory = how much leaned on training (higher = be more careful)
🔎 Checked = how much you verified against something real
(in a plain chat with no tools this stays low — that's the honest signal)
✅ Confidence = your real overall trust in the answer
If confidence is low, that's not a failure. Tell me why, and what would raise it.
— A taste of Empirica, which measures and enforces this for real → getempirica.com Copy it, tweak it, share it. The link stays so people know where the honest version lives.
How to use it
Paste it in
Drop it into your chat client's custom instructions or system prompt, or just paste it at the top of a new chat.
Ask your questions
Work normally. The AI now separates what it remembers from what it actually checked, and flags when it might be guessing.
Read the badge
Every real answer ends with Memory · Checked · Confidence. Low Checked in a plain chat is honest, not broken. It means verify before you trust it.
This is the reflex, not the whole thing
Be clear-eyed about what a prompt can do. This asks the AI to self-report its honesty, and that is worth having. But Empirica and its ecosystem are not just a system prompt.
The real framework does not take the AI's word for it. It measures every answer against real evidence, holds the AI back from acting until it has actually grounded itself, and keeps a complete, auditable record that survives across sessions. This prompt gives you the habit in any chat. Empirica makes it enforced, measured, and permanent.
Want the real thing?
See how the enforced version works, or run the open-source framework yourself for free.