What it is
An instruction for the gaps. On the ten checks this is Doubt, the ninth check, and it is the one ordinary prompts miss most completely: in the thinkhard.si study of 319 one-line work requests, not one said what to do when something was missing.
There are three useful instructions, and the right one depends on where the answer is going.
- Ask first, when you are there to answer and the gap would change the whole draft, such as a bio that needs your current title.
- Flag it in place, with a marker such as
[CHECK], when you will review the draft anyway, as with minutes or a contract summary. - Leave it blank and say so, with a fixed phrase such as “not stated”, when the answer feeds a table or a system, where a guess would be counted as data.
When it helps
It helps wherever a wrong detail costs something: numbers, names, dates, owners, prices, legal terms. A summary that invents an owner for an action item sends someone a task they never agreed to. An extraction that fills in a typical notice period makes a contract look safer than it is.
It helps most with prompts that run on inputs you will not read closely, such as a saved prompt the team uses on every contract or every ticket. Each input has different gaps, and a rule for gaps handles all of them the same way.
When it does not
For brainstorming and drafting from scratch, guessing is the job. Asking for flags on a list of campaign ideas just gets you flags.
If every answer comes back full of flags, the problem is the material, not the instruction. Give it the facts it keeps asking about, and the flags will drop to the real gaps.
“Ask me first” needs someone to ask. In a prompt that runs without you, such as an automation or a batch of files, there is nobody to answer, so use a flag or a fixed phrase instead.
“Ask if anything is unclear” on a simple task can produce a round of questions you did not need. Name the gaps that matter, such as the date or the owner, so the model asks about those and makes sensible choices about the rest.
And it reduces guessing; it does not end it. A model can still fill a gap without flagging it, so check the facts that matter whether or not they carry a flag. The Fact Check List pulls every number, date and quote out of an answer so you can verify each one.
Before and after
Three requests where the after version says what to do with a gap. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
From the contract below, list the term, the renewal terms, the notice period and the fees. """ [the contract, pasted in full] """
From the contract below, list the term, the renewal terms, the notice period and the fees. If an item is not in the contract, write "not stated". If the wording is unclear, quote the clause and mark it [CHECK]. Do not fill a gap from what similar contracts usually say. """ [the contract, pasted in full] """
A contract with no notice clause can come back from the first version with a notice period anyway, the one contracts like it usually have. The second version says “not stated”, which is the most important thing you could learn about that contract, and quotes the clauses you need a lawyer to read.
Write a 60-word speaker bio for me for the conference program, from my profile below. """ [your profile, pasted in full] """
Write a 60-word speaker bio for me for the conference program, from my profile below. If the bio needs something the profile does not give, such as my current title or where I'm based, ask me before you write it. """ [your profile, pasted in full] """
With an old profile, the first version writes the bio around your last job title. The second asks for the current one before writing, which is one short reply instead of a correction the conference prints.
Write next quarter's sales forecast summary from the pipeline export below. """ [the pipeline export] """
Write next quarter's sales forecast summary from the pipeline export below. If a deal has no close date or no amount, leave it out of the totals and list it under "Needs an update" instead of guessing. """ [the pipeline export] """
Most pipeline exports have a few incomplete deals. The first version decides for itself whether to count them; the second keeps them out of the totals and hands sales a list of records to fix.
Each after version passes Doubt and, because the rule sits in its own paragraph, Fit; the contract version also passes Limits for its last sentence.
The line to add
Pick the version that suits where the answer is going:
If {{the_fact_that_matters}} is missing or unclear, ask me before you write.
If anything is unclear, write [CHECK] in its place and list every [CHECK] at the end.
If an item is not in the material, write "not stated". Do not fill it in from general knowledge.
The list at the end of the second version matters. A flag in the middle of a page is easy to miss; a list of every flag is a checklist.
In ChatGPT, Claude, Gemini and Copilot
All four vendors point the same way. Anthropic’s guide to reducing hallucinations puts first “allow Claude to say I don’t know”: give it explicit permission to admit uncertainty, which it says can drastically reduce false information. The same page notes that these techniques reduce hallucinations without eliminating them.
OpenAI’s guide to its latest models, the ones behind ChatGPT, describes its newest model as filling routine gaps from context and asking focused questions when the answer could change the outcome, and advises telling the model when an important ambiguity should trigger a question. That is the case for naming the gaps that matter. Google’s guide for Gemini suggests a grounding instruction for answers that must come from supplied material: if the exact answer is not written there, say the information is not available.
Microsoft’s guide to Copilot prompts tells you to review and verify the responses you get, and to expect some back-and-forth. A flag tells you where to look first.
Key takeaways
- Without a rule for gaps, a model fills them with plausible guesses.
- Ask first when you are there, flag when you will review, write “not stated” when the answer feeds a table.
- Name the gaps that matter so the model does not ask about trivia.
- Many flags mean thin material. Add the facts.
- Flags reduce guessing without ending it. Check what matters anyway.
Questions
How do I stop ChatGPT from making things up?
Give it the facts, tell it to use only those, and tell it what to do when one is missing: ask you, mark the gap [CHECK], or write “not stated”. That last instruction is the one most prompts leave out. It reduces invented details without removing them, so check anything that matters. The guide to reducing hallucinations goes further.
How do I get AI to say it does not know?
Say that it may, and what to write instead: “If the answer is not in the document, say so rather than answering from general knowledge.” Anthropic’s guidance calls this giving the model permission to admit uncertainty, and says it can sharply reduce false information. Without that permission, the model is more likely to produce an answer anyway.
Should AI ask me questions before it answers?
When a missing fact would change the answer, yes. Name the facts that matter so it does not ask about trivia. For a bigger job, the questions-first pattern gets all the questions in one round before any drafting starts.