How to Make Claude More Critical of Your Ideas

To make Claude more critical, ask it to review a specific proposal against explicit criteria, challenge its assumptions, and support each objection with evidence or a test. Tell it that finding no serious problem is an acceptable outcome. You want a useful assessment, not a performance of disapproval.

Keep the eventual choice separate from the critique. fenbs can preserve the decision a person accepts and the reasons behind it, so a later assistant does not confuse a rejected objection with an instruction. The review itself can begin with one proposal and the prompt below.

What should you ask Claude to do first?

1. State the decision: name the proposal, the alternatives, and what you are trying to achieve. Supply the actual draft, design, or plan rather than only your enthusiastic summary.
2. Set three to five criteria: describe what a good result must accomplish and which constraints cannot change. Separate requirements from preferences.
3. Request the strongest objections: ask which assumptions could break the proposal, how serious each problem would be, and what evidence supports that concern.
4. Ask for repairs and tests: each major objection should lead to a possible improvement or a practical way to investigate it.
5. Make the decision yourself: accept, reject, or defer the proposed changes, then record why. Claude should not silently turn its review into permission to act.

Give context that changes the judgment: who will use the result, what has already been tried, and what success means. Anthropic recommends clear, explicit instructions↗ (opens in a new tab). A defined review is easier to assess than a request to “be tougher.”

Which review criteria make criticism useful?

Choose criteria before reading the verdict. Otherwise, it is easy to accept whichever standard supports the answer you already want. A proposal may be attractive under one standard and weak under another; the review should make that tradeoff visible.

For a small design studio, a landing page might need to explain the service, make the next action obvious, and remain readable on a phone. “I prefer a quieter layout” is a preference. “The contact action is absent on the supplied phone screen” is an observation the designer can check.

Anthropic’s evaluation guidance↗ (opens in a new tab) recommends specific, measurable success criteria. You can apply that principle without building a scoring system: write down what would count as passing each important requirement.

How can a prompt produce usable objections?

Review the proposal below before I decide whether to proceed.
Goal: [desired result]
Audience: [who uses it]
Criteria: [three to five requirements]
Constraints: [fixed limits]

Identify the strongest objections, not a required number of flaws. For each: give the affected criterion, the evidence, the assumption involved, the consequence, and a repair or test. Distinguish observed problems from possible risks. Keep strengths that matter. Say when the evidence is insufficient. Do not revise or execute anything yet.

Proposal and supporting material: [paste here]

The distinction between a problem and a risk matters. An inconsistent date in a draft is present in the material. The possibility that readers will dislike its tone is a hypothesis. Both may deserve attention, but they need different checks and should not arrive with the same certainty.

Ask for the strongest counterargument to the preferred approach, then the strongest reply to that counterargument. This makes the review examine whether a concern is decisive, already addressed, or simply a tradeoff you knowingly accept. Do not require equal space for every alternative.

If you are new to the product, what Claude is provides the broader background. For evaluating visual work specifically, use the examples in Claude design alongside your actual design requirements.

What does a useful critical review look like?

Suppose you propose removing a project intake form to make inquiries easier. “This is risky” does not help. A useful objection says that removing the form also removes the project type your team uses to route inquiries, then asks whether the replacement captures that information another way.

A practical repair might be one optional project-type question, or a short follow-up message. A useful test would compare whether sample inquiries can still be routed. This example does not establish which option is best; it shows what makes a criticism actionable.

Ask Claude to rank concerns by their effect on the goal. A missing requirement should receive attention before a minor stylistic preference. Avoid treating a long list as a better review: five versions of the same objection do not represent five independent reasons to change course.

How do you check the critique before changing anything?

Inspect the most consequential objections first. Open the cited source, compare the quoted passage with the claim, or reproduce the stated problem. If Claude claims that a sentence is missing, find the relevant part of the draft yourself. An authoritative tone is not supporting evidence.

Anthropic’s hallucination guidance↗ (opens in a new tab) recommends allowing uncertainty and checking claims against sources. Give Claude an explicit way to say “not enough information,” especially when the proposal depends on facts you did not supply.

Save accepted changes separately from concerns that remain unproven. A person might accept the routing fix but reject the suggested tone change because customer examples support the current wording. Decisions and rules in fenbs keep that accepted choice available for later work.

How to do this with fenbs

1. Create a free account, then add the project and a task describing the proposal you need to review.
2. Record any continuing review requirements in Decisions and rules, with the person who agreed to them. Keep one-time evaluation criteria with the relevant work.
3. Connect Claude through the guide in Integrations. Ask it to read the project context before reviewing the supplied proposal.
4. After you choose a direction, ask the assistant to record your decision and its reasons. Save unresolved concerns and the next check with the task, then save progress before ending the session.

fenbs fits when the critique affects ongoing work or several assistants need the same accepted answer. For a single paragraph you will never revisit, a direct review in Claude may be enough; fenbs does not judge whether an objection is correct.

When should you stop asking for more criticism?

Stop when the remaining concerns are low impact, cannot be resolved with the available evidence, or need a real-world test. More generated objections can keep a reasonable decision suspended indefinitely. Set a stopping condition, such as resolving every unmet requirement and assigning the remaining material risks a test.

For separating evidence from assumptions across competing explanations, see making Claude more objective. The wider AI prompts collection covers other ways to shape a response. Here, the goal is a stronger proposal and a defensible human choice.