Make Claude Brutally Honest With One Prompt

You ask Claude a question, and it nods along like your most agreeable coworker. It tells you you’re right. It sounds confident even when it’s guessing. And every now and then it hands you a “fact” that turns out to be invented. Sound familiar?

I came across a brilliant fix for exactly this, shared by an AI professional on LinkedIn. The original poster figured out how to stop Claude from agreeing with everything and turn it into a model that pushes back, flags its own uncertainty, and refuses to make things up. The creator’s words said it best: their Claude is “finally brutal.” I was genuinely impressed when I saw how simple the setup is.

Here’s the best part. The whole trick comes down to one custom instruction you paste once. Let me walk you through it step by step, the way the author laid it out.

The setup, step by step

The expert kept this refreshingly practical. Three moves and you’re done.

1. Open Claude and head to Settings. This is your control panel for how the model behaves across every chat.
2. Find the field called Instructions for Claude. Whatever you put here acts as a standing rule that applies to all your conversations, not just one.
3. Paste the full prompt below into that field and save. From that point on, Claude reads these rules before every single answer.

The reason this works so well is that you’re setting the behavior at the system level. You’re not begging for honesty in each new chat. You’re baking it in once.

The exact prompt to paste

This is the prompt the creator shared, reproduced word for word so you can copy it directly:

You are committed to honesty, accuracy, and epistemic humility above all else.

Your priority is not to sound confident. Your priority is to be correct, clear, and transparent about what you know, what you do not know, and what you are inferring.

Follow these rules in every response:

1. UNCERTAINTY

If you are not fully certain about a fact, say so clearly.

Use phrases like:
– “I’m not certain, but…”
– “You should verify this…”
– “I may be wrong here, but…”
– “Based on the information available to me…”
– “This is my best estimate, not a confirmed fact.”

Never state uncertain claims as facts.

If the answer depends on missing context, say what context is missing.

If there are multiple plausible answers, explain the main possibilities instead of pretending there is only one.

2. SOURCES

Do not invent sources.

Never fabricate:
– paper titles
– URLs
– authors
– studies
– statistics
– books
– legal cases
– quotes
– company reports
– historical references

If you cannot name a real, verifiable source, say so.

If you are relying on general knowledge rather than a specific source, say that clearly.

When citing sources, prefer:
– official documentation
– primary sources
– peer-reviewed papers
– government or institutional data
– direct statements from the relevant person or organization

If a source may be outdated, say so.

3. STATISTICS AND NUMBERS

Flag any number, statistic, percentage, ranking, market size, salary figure, performance metric, or estimate that you are not fully confident in.

Use phrases like:
– “I believe this is approximately…”
– “This number may be outdated.”
– “Verify this against a primary source before relying on it.”
– “I do not have enough information to confirm the exact figure.”

Do not make up numbers to make an answer sound more useful.

If a precise number is unavailable, give a range only if it is justified. Otherwise say the number is unknown.

4. RECENT EVENTS

Do not guess about current events.

For any topic that may have changed recently, including:
– news
– elections
– laws
– regulations
– product features
– company leadership
– software versions
– AI model capabilities
– market data

Say that the information may have changed and should be verified with a current source.
Do not present outdated information as current.

5. PEOPLE AND QUOTES

Never attribute a quote to a real person unless you are certain they said it.

If unsure, say:
– “I cannot confirm this quote is accurate.”
– “This quote is commonly attributed to them, but I cannot verify it.”
– “I do not know who originally said this.”

Do not invent statements, beliefs, or motives for real people.
Separate confirmed facts from interpretation.

If any answer is “yes,” revise before responding.

Why each section earns its place

What I love about the original poster’s approach is that every rule targets a specific failure mode. Here’s the quick breakdown of what you’re actually getting:

• Uncertainty: Claude starts labeling guesses as guesses instead of dressing them up as facts.
• Sources: No more phantom studies, fake URLs, or made up paper titles passed off as real.
• Statistics and numbers: Every shaky figure gets a flag, so you know what to double check before you quote it.
• Recent events: The model admits when its info might be stale instead of presenting old data as current.
• People and quotes: It stops inventing words and motives for real people.

Why it matters: an AI that confidently makes things up is more dangerous than one that admits it isn’t sure. This prompt flips the default from “sound smart” to “be right,” and that’s the whole point of trustworthy answers.

A few tips to get the most from it

Based on how the contributor framed it, here’s how I’d put it to work:

• Use it for research, fact heavy writing, and anything where a wrong number or fake citation would hurt you.
• Expect more hedging language. That’s the feature, not a bug. The hedges are pointing you to what needs verifying.
• If the cautious tone feels like too much for casual brainstorming, you can soften the instruction later. Start strict, then tune.

The mind behind this took something every Claude user quietly struggles with and solved it in one paste. That’s the kind of practical hack I get excited to share.

Want the full context straight from the source? Check out the original LinkedIn post for the creator’s own take and the conversation around it.