An instruction to show the evidence can make an AI answer easier to check. The checking is still our job.
“Does this university require predicted IB grades?” sounds like a simple question. It isn’t. Does the answer apply to this admissions cycle? To applicants completing the full IB Diploma or also to students taking individual IB courses? Do the grades go in with the application, or does the university request them later?
An AI assistant can give you a tidy answer to the first question without ever addressing the other three. And if the answer arrives in a neat table with a university link, it’s surprisingly easy to mistake that neatness for certainty.
A LinkedIn post from Harish Kumar offers a useful starting point. He suggests giving Claude standing instructions to flag uncertainty, avoid invented sources, and ask when context is missing. I’d stop short of his claim that a prompt can make Claude stop making things up. Anthropic’s Claude Opus 4.6 system card identifies invented facts and citations as a risk observed in external pilot use. That documents the problem; it doesn’t establish whether this particular instruction reduces it. What I want from the instruction is an answer I can check quickly.
A standing instruction for college counseling
Here’s the version I’d use:
When I ask about a college policy, deadline, cost, or eligibility rule, tell me which admissions cycle and applicant group your answer applies to. If that information is missing and could change the answer, ask me for it.
For each claim that affects what a student or counselor needs to do, give me the official source you actually checked and point me to the relevant section. Distinguish what the source explicitly says from what you infer. If the source doesn’t answer the question, or if two official sources disagree, say what remains unresolved. Never invent a source or imply that you opened a page you couldn’t access.
End by identifying anything I still need to verify before passing the answer along.
In Claude, Anthropic’s instructions point to Settings → Your Claude → Instructions. In the older layout, look under Settings → Account → Instructions for Claude. Instructions saved here apply across your conversations, so you may prefer to paste this into an individual chat if you only want it for policy research.
What this looks like in practice
Back to those predicted grades. Instead of asking, “Does this university need them?”, try: “For first-year applicants completing the full IB Diploma outside the U.S., what does the university say about submitting predicted grades for fall 2027? Who submits them, and by when? If the published guidance doesn’t say, identify what I need to ask admissions.”
Suppose the assistant finds a page that requires a school transcript but says nothing about predicted grades. That’s useful information. It tells me where the published guidance stops and gives me a more precise question for the admissions office. It doesn’t tell me that predicted grades are optional.
I’d test the instruction on a policy you already know well. Ask the same question with and without it. Open the cited page both times. Can you find the sentence that supports the answer? Does it apply to the student you’re advising? Did the assistant flag what the page left out?
The goal is to spend less time hunting for the right page and more time thinking about what it means for the student. If the instruction helps me do that, it’s worth keeping.