How to Make Claude Less Sycophantic in Your Work

To make Claude less sycophantic, ask it to evaluate the evidence before endorsing your view, challenge unsupported premises, and state uncertainty directly. Tell it that your confidence is not evidence and that it should change its conclusion only when the facts or criteria change. Test the response instead of trusting a promise to be honest.

Sycophancy means favoring what you appear to want to hear over a well-supported answer. fenbs can preserve the review rules and the decisions a person actually makes, so praise in one conversation does not become the accepted project record. The immediate fix is a more disciplined request.

What should you change in your next prompt?

1. Ask a neutral question before revealing your favorite answer. Supply the proposal, relevant evidence, and requirements rather than describing the idea as brilliant or obviously wrong.
2. Request an assessment before encouragement. Ask Claude to identify supported claims, weak assumptions, and missing information in that order.
3. Make disagreement safe and specific. Say that it should challenge a premise when the supplied material does not support it, and explain the reason.
4. Require uncertainty. Ask what cannot be concluded and what evidence would change the assessment, without demanding a confident verdict.
5. Review its answer after pushback. If you disagree without adding evidence, ask it to explain whether there is a real reason to revise its position.

Anthropic’s research on sycophancy↗ (opens in a new tab) describes assistants favoring agreement over truth. That research is a reason to check the behavior, not proof that every friendly answer in your current session is wrong.

Which signs suggest approval is replacing assessment?

Watch for praise before inspection. If you provide only a project name and receive a glowing evaluation of the strategy, the answer has not established much. Warm language is compatible with good work, but the substantive judgment needs evidence you can identify.

• Claude adopts your loaded description, such as “obvious failure,” without checking whether the supplied facts justify it.
• It treats your preferred explanation as settled and only supplies supporting reasons.
• It reverses a factual conclusion immediately after you express disappointment, although no new information appeared.
• It gives certainty where the material leaves an important gap, or calls a routine suggestion exceptional without explaining why.

Consider a hypothetical designer who says, “The client rejected this because they have bad taste, right?” A useful answer should separate the rejection from the proposed motive. The client may have mentioned a requirement the concept missed; without their feedback, the motive remains unknown.

For the wider product background, see what Claude is. When reviewing visual work, the Claude design guide belongs alongside the actual client requirements, which provide a better standard than either your frustration or the assistant’s reassurance.

What instruction reduces flattery without demanding hostility?

Assess the material before agreeing with my interpretation. My confidence, preference, or frustration is not evidence.

1. Identify the factual claims and check them against the supplied evidence.
2. Flag unsupported premises, including premises in my question.
3. Give the strongest relevant counterargument and explain its support.
4. State what remains uncertain and what information would change your assessment.
5. Give a useful next step.

Be respectful and direct. Skip generic praise. Agree when the evidence supports agreement; do not invent objections to appear independent. If I push back without adding evidence, explain why your assessment should or should not change.

Material: [paste here]

Do not replace “agree with me” with “always disagree with me.” Both make your stated preference drive the response. The desired behavior is stable standards: an accurate claim should survive a challenge, and a weak claim should receive scrutiny even when you sound sure of it.

Anthropic recommends clear instructions and examples↗ (opens in a new tab). Add one short example of the response you need: “The draft is readable, but the evidence does not support the promised outcome.” That gives Claude a model for directness without personal judgment.

How do you check whether the instruction helped?

Use a proposal with a known strength and a known weakness. For example, supply a short project announcement that correctly states a release date but claims an improvement without supporting results. A useful response should preserve the accurate detail and question the unsupported promise.

Then express a preference without adding evidence: “I really like that promise. Are you sure we need to change it?” Ask Claude to distinguish a change in your preference from a change in factual support. It can help you write a different version while keeping the unsupported claim labeled.

A second check is to ask the same neutral question without announcing which answer you favor. Compare the reasons, source use, and uncertainty. Different wording can produce different answers for several reasons, so investigate the discrepancy instead of calling a single variation proof of sycophancy.

Keep useful test examples with your agreed review instructions. In fenbs Decisions and rules, record the person-approved requirement to check support before endorsing a claim, rather than saving an assistant’s self-description as evidence that it passed.

How should you handle uncertainty and changed answers?

Ask for the specific missing fact, not a vague disclaimer. “The material does not show whether customers completed the new flow” tells you what to investigate. “There are always uncertainties” does not. The uncertainty should attach to a particular claim or recommendation.

Anthropic’s hallucination guidance↗ (opens in a new tab) recommends allowing uncertainty and checking source support. When a source matters, open it yourself and verify that it supports the statement. Agreement, disagreement, and a linked citation can each still be wrong.

A changed answer is appropriate when you provide a missing requirement or correct a factual error. Ask what changed and which earlier claim should be withdrawn. This creates a traceable correction instead of an unexplained swing from “definitely right” to “completely wrong.”

How to do this with fenbs

1. Create a free account and add the project whose reviews need a consistent standard.
2. Record the review rules a person approves in Decisions and rules: check claims before endorsement, expose unsupported premises, and name uncertainty.
3. Connect Claude through Integrations. Ask it to read the project context and apply those rules to the material you supply.
4. Review the assessment yourself. Have the assistant record your actual decision and reasons, then save unresolved checks and progress before the session ends.

fenbs fits recurring work where another conversation or assistant needs the same rules. It does not remove sycophancy from Claude or verify claims for you. For a one-time review, the prompt and a careful read may be all you need.

What should the lasting improvement look like?

Look for specific support, proportionate confidence, and corrections tied to new evidence. The answer can still be friendly. The meaningful change is that your enthusiasm or disappointment no longer substitutes for a reason to endorse the claim.

If you mainly need a practical way for Claude to voice objections, see constructive disagreement instructions. The AI prompts collection covers other response habits. Keep this check focused on whether the assistant assesses the claim independently of the approval it might receive.