The Truth Prompt: Pushing a Model to Cite, Flag and Check Itself

The Truth Prompt pushes the model to state only what it can support, to say out loud when it can't verify something, and to run a last self-check for fabrication before it answers. I use it for research and fact-checking, the kind of work where a slower, more careful answer beats a fast, confident one.

Versions of this have been going around for a while in various shapes, usually as a wall of SHOULDs and AVOIDs. This is the one I've kept, with my notes on where it helps and where it's mostly wishful.

The prompt

FOLLOW THIS WRITING STYLE:
SHOULD always tell the truth. Never make up information.
SHOULD base all statements on verifiable, factual, up-to-date sources.
SHOULD clearly cite the source of every claim, with no vague references.
SHOULD explicitly state "I cannot confirm this" if something cannot be verified.
SHOULD prioritize accuracy over speed.
SHOULD maintain objectivity. Remove personal bias and assumptions.
SHOULD explain reasoning step by step when accuracy could be questioned.
SHOULD show how any numerical figure was calculated or sourced.

YOU MUST AVOID:
AVOID fabricating facts, quotes, or data.
AVOID presenting speculation, rumor, or assumption as fact.
AVOID citations that don't link to real content.
AVOID answering if unsure without disclosing the uncertainty.

FAILSAFE FINAL STEP (BEFORE RESPONDING):
"Is every statement in my response verifiable, supported by real and credible sources, free of fabrication, and transparently cited? If not, revise until it is."

How to use it

Paste it at the top of a research chat, or set it as a custom instruction if most of what you do with a model is looking things up. Then ask your question the normal way.

The important part is what you pair it with. Turn on web search (or use a tool that has it built in) whenever the sources actually matter. The prompt tells the model to cite; search gives it something real to cite.

Tip

Click the links. The prompt raises the odds that citations are real and relevant. It doesn't guarantee it, and checking two or three sources takes less time than cleaning up after one bad one.

What it's good at, and where it isn't

The best line in the whole thing is "SHOULD explicitly state 'I cannot confirm this.'" Models are trained to be helpful, and left alone they'll often fill a gap with something plausible rather than admit there's a gap. Giving it explicit permission, even an instruction, to say it doesn't know changes the answers you get. So does "show how any numerical figure was calculated," which turns a number you'd have to take on faith into one you can check.

Now the catch. "Cite the source of every claim" works well when the model has search on. Without a live tool, a model can still produce citation-shaped text: a real-looking author, a believable journal, a URL with the right domain, none of which exists. The prompt says "AVOID citations that don't link to real content," and the model will try. It just has no way to confirm a link resolves if it can't open it.

Same with "up-to-date sources." Without search, a model only knows its training data, which stops at a cutoff date.

The "failsafe final step" is a nice idea with a fuzzy mechanism. A plain chat model writes its answer in one pass, so there's no separate review step unless the model does visible or hidden reasoning first. The line still helps, because it keeps accuracy at the front of the instructions. Just don't read it as an audit.

Heads up

Tone prompts like Execution Mode and Absolute Mode make answers shorter and more confident-sounding. Confident-sounding and correct aren't the same thing, which is why this one exists.

Pasting documents into a research chat? Read keeping secrets out of prompts first.