The Anti-Hallucination Stack: 3 Claude Instructions

Claude, like every large language model, is trained to produce fluent, confident text - and that training does not distinguish between facts it knows and facts it is guessing. The output sounds identical either way. One wrong stat in a proposal can cost a deal; one fabricated quote in a post can dent credibility you spent months building. Most people fix this by checking everything manually, which erases the time savings that made AI worth using. The smarter fix is to change how Claude handles uncertainty before it writes a word.

By default Claude states uncertain things as if they are fact. This instruction tells it to label its confidence on any claim that could be wrong. When you see 'I believe' or 'you may want to verify this', you know which lines to check and which to trust - so you spot-check the flagged lines instead of auditing the whole document.

Instruction 1 - uncertainty flagWhenever you are not fully confident in a fact, statistic, date, name or
claim, explicitly flag it. Use phrases like 'I believe', 'I'm not
certain but', or 'you may want to verify this'. Do not present uncertain
information as established fact. If you do not know something, say so
directly rather than guessing.

This forces source transparency. Without it, Claude blends three things into one seamless answer: information from your documents, information from its training data, and its own inferences. With it, every major claim is tagged with its origin, so you know immediately which claims are solid and which need a source check.

Instruction 2 - source transparencyWhen you include facts, statistics, research findings or specific claims,
indicate the basis for each one using one of three labels: (1) 'Based on
the documents you provided:' for my materials, (2) 'From my training
data:' for your knowledge base, or (3) 'This is my inference:' for
conclusions you are drawing. Apply these labels consistently.

The most important one. It addresses the root cause: Claude's tendency to complete a task even when it lacks the information to do it accurately. Ask for a case study with specific numbers it does not have, and by default it will generate plausible-sounding numbers to be helpful. This instruction gives it explicit permission to stop and ask instead - reframing an incomplete answer as the correct behaviour when real information is missing.

Instruction 3 - refusal over fabricationIf you do not have sufficient accurate information to complete a task or
answer a question, do not guess or fabricate details to fill the gap.
Instead, tell me specifically what information you are missing and ask me
to provide it. An incomplete response that asks the right follow-up
questions is more useful than a complete one built on invented
information. Accuracy is more important than completeness.

Each works on its own; combined they create a system where Claude flags uncertainty, tells you where every claim came from, and asks for missing information instead of inventing it. Place all three at the top of your system prompt (or in your CLAUDE.md) so they apply to every message automatically - set once, running in the background on every output. They are plain text and work in Claude.ai, in the API, and in any tool built on Claude.