Two sources, one confident voice
When you paste a report and ask a question about it, the answer can come from two very different places. It can come from your document. Or it can come from everything the assistant absorbed about reports like yours during training, which is a great deal, most of it plausible, none of it yours.
The problem is that both arrive in the same voice. The sentence built from your page eleven and the sentence built from a general sense of how these documents usually go are indistinguishable on the page. Same fluency, same confidence, same clean formatting. This is where fabricated statistics and invented citations come from: not from the assistant lying, but from it filling a gap in the most convincing available style.
Grounding is the practice of forcing the answer to come from one place, and making it show you which sentences came from where.
A confident answer is not a sourced answer. Fluency is produced by the same machinery whether or not there is anything behind the claim, so it can never be your evidence that there is.
Why "work only from this" changes behaviour
It sounds like a polite request. It is actually a change of task.
Without it, the task is broadly "produce a good answer about this subject," and your document is one input among many. The assistant is free to smooth over a gap, because a smooth answer is a good answer under that instruction.
With it, the task becomes "extract and report what this specific text says." Now a gap in your document is not a problem to be smoothed. It is a finding to be reported. That is the shift: gaps stop being embarrassing and start being output.
You have to make that explicit, though, because the default is to be helpful. Left to itself, Claude or Gemini will bridge a missing piece rather than announce it, since announcing it feels unhelpful. Tell it that reporting the gap is the helpful behaviour, and it will do that instead.
In plain English
- Grounding:
- Requiring an answer to be built only from source material you supplied, rather than from the model's general knowledge.
- Fabrication:
- A confident, plausible claim with nothing behind it, produced when the model fills a gap instead of flagging it.
- Inference:
- A reasonable conclusion drawn from the source that the source does not actually state.
- Quote requirement:
- Making the assistant paste the exact words it relied on, so a claim can be checked against the source in seconds.
The instruction worth saving
This is the reusable block. Keep it somewhere you can paste it, because you will use it constantly.
GROUNDING RULES for everything that follows.
Use only the material I have pasted below. Do not use general knowledge,
do not use anything from earlier in this conversation, and do not use
your own previous answers as a source.
Label every sentence in your answer with one of these:
[STATED] the source says this. Follow it with a short exact quote.
[INFERRED] a reasonable conclusion the source does not actually state.
Say which lines you inferred it from.
[MISSING] the source does not settle this point.
Rules:
- Never replace a MISSING with a plausible estimate.
- If the source contradicts itself, quote both parts and mark it CONFLICT.
- If you cannot answer at all from this material, say so and stop.
- A short answer with three labelled sentences is better than a long one
with any unlabelled sentence.
Material:
[paste your source here]
That last rule matters more than it looks. Without it, a length instinct takes over and the labels start decorating padding. Say plainly that a short grounded answer beats a long ungrounded one, and you remove the incentive to fill.
The quote requirement does the heavy lifting
Of all of it, the exact quote is the piece that changes outcomes most. A claim with a quote attached can be verified with a search across your own document in about four seconds. Either the words are there or they are not.
It also acts as a brake before the fact. A sentence that has to arrive with its evidence attached is harder to produce loosely, because the requirement is checkable in a way that "be accurate" is not.
โ Weak prompt
Prompt
[pastes an internal research report] What did the study find about retention?
Output
The study found a meaningful improvement in retention following the intervention, with participants in the treatment group showing roughly a 20 percent uplift over the control group. Findings were consistent across the sample and align with what similar programmes typically report.
Read that last clause again. It compares your study to what similar programmes typically report, which is not in your document at all. The 20 percent has no quote behind it, so you cannot tell whether it came from page nine or from thin air.
โ Good prompt
Prompt
Use only the report pasted below. Label every sentence STATED with an exact quote, INFERRED with the lines it came from, or MISSING. Never estimate a number. What did the study find about retention?
Output
[STATED] Retention at six months was 68 percent in the treatment group and 61 percent in the control group. Quote: 'six month retention was 68 percent versus 61 percent'. [INFERRED] That is a 7 point difference, calculated from the two figures above. [MISSING] The report does not state whether this difference was statistically significant, and gives no confidence interval or p value.
Every number is now attached to words you can find in your own file. The comparison to other programmes vanished, because it was never in the source. And the missing significance test is now a visible finding instead of a silent hole.
Two follow-ups worth running
Grounding on the way in is most of the job. These two catch what slips through.
Go back through your last answer sentence by sentence.
For each sentence, paste the exact words from my source that support it.
If you cannot find supporting words, say NO SOURCE and delete the
sentence from a revised version.
Give me the revised answer with only the sentences that survived.
Based only on the material I supplied, list the questions a careful
reader would ask that this material cannot answer.
Do not answer them. Do not speculate about what the answers might be.
Just list the gaps, and note which one would matter most.
That second prompt is the one people skip and then wish they had run. It turns the silent part of your evidence base into a list you can act on, which is usually more valuable than another paragraph of summary.
Deleting unsupported sentences beats asking for them to be rewritten. Rewriting produces a softer version of the same unsourced claim, complete with a hedge that makes it sound carefully considered.
Checkpoint
Grounding means the answer comes only from material you supplied, with claims labelled as stated, inferred or missing, and exact quotes attached. Saying work only from this changes the task from producing a good answer to reporting what your text says.
๐ Quiz
Question 1 of 4Why can you not tell a sourced claim from an unsourced one by reading the answer?