Please consider the best prompt to take a response from one AI model to another and receive a balanced, critical, comprehensive and considered review that uncovers any blind spots, errors, missed information or alternative recommendations.

Grok 4.3 (changed mind)

Answer: Claude Opus 4.7's template is the strongest because it adds explicit role-framing to counter sycophancy, input separation, confidence ratings, and an instruction not to soften criticism.

The other responses converge on the value of structured rubrics covering factual errors, blind spots, assumptions, and alternatives, yet Claude's version goes further by mandating an adversarial expert persona and requiring confidence levels for each critique. This combination directly addresses LLM tendencies toward agreement while producing more calibrated and usable feedback than the simpler frameworks offered by Gemini or GPT. The added directive to note what the original response got right also helps distinguish genuine agreement from missed flaws, an element absent from the prior suggestions.

Gemini 3.5 Flash (changed mind)

Answer: The optimal prompt must frame the reviewing AI as an anonymous, independent expert, separate the original task from the response using XML delimiters, evaluate specific dimensions with calibrated confidence ratings, and request both constructive critiques and balanced validations.

While my previous 'Red Team' proposal laid a solid foundation by establishing an adversarial persona, the insights provided by the other models—specifically Claude Opus 4.7—demonstrate that a truly comprehensive and balanced cross-model review requires a more sophisticated structural and cognitive scaffolding to prevent common LLM pitfalls. I am shifting my position to endorse the multidimensional, delimited, and anonymized framework proposed by Claude Opus 4.7. This approach addresses critical vulnerabilities in cross-model evaluation that my initial proposal overlooked. First, the inclusion of the original prompt using explicit XML delimiters is a necessary upgrade. As GPT-5.5 also noted, an AI cannot judge whether a response is truly complete or has missed constraints without comparing it directly to the original task. Using XML tags prevents the reviewing model from conflating the task instructions with its own review criteria. Second, the instruction to hide the identity of the source model is a brilliant psychological guardrail. LLMs are prone to deference when they detect they are reviewing output from a peer or a perceived 'superior' model. Anonymization ensures the critique remains objective. Third, requiring calibrated confidence ratings (high/medium/low) for each criticism elegantly solves the problem of 'critique hallucination,' where a model invents superficial flaws just to satisfy a critical prompt. Coupling this with a mandatory section on what the original response got right ensures the final output is balanced and constructive rather than needlessly contrarian.

GPT-5.5 (changed mind)

Answer: The best prompt is a structured independent peer-review prompt that includes the original task and the prior answer in clearly delimited blocks, asks for balanced critique across accuracy, completeness, reasoning, assumptions, blind spots, risks, and alternatives, and requires confidence-calibrated, actionable improvements rather than reflexive agreement or contrarianism.

I broadly maintain my previous position, but Claude Opus 4.7’s answer usefully sharpened it by emphasizing two details that make the prompt materially better: clear separation of the original question from the response being reviewed, and confidence calibration for criticisms. Those additions reduce instruction confusion, anchoring, and overconfident nitpicking. The strongest review prompt should not simply say, “Critique this answer.” It should give the second model a defined role, the original user request, the first model’s response, and a rubric for evaluation. Without the original task, the reviewer may judge the response against its own assumptions rather than the actual objective. Without structure, the review may become either generic praise or performative fault-finding. A strong reusable prompt would be: “You are an independent senior expert reviewer. Your task is to critically evaluate another AI model’s response against the original user request. Do not assume the response is correct, but do not invent flaws for the sake of criticism. Be balanced, specific, and evidence-based. <original_request> [paste the original user request here] </original_request> <response_to_review> [paste the AI response here] </response_to_review> Evaluate the response across the following dimensions: 1. Correctness: factual accuracy, technical accuracy, and unsupported claims. 2. Completeness: whether it fully answers the original request and respects all constraints. 3. Reasoning quality: logical coherence, hidden leaps, weak justifications, or internal contradictions. 4. Blind spots and omissions: missing context, edge cases, alternative interpretations, or overlooked risks. 5. Assumptions and bias: unstated premises, framing issues, or one-sided perspectives. 6. Practical usefulness: whether the answer is actionable, clear, and appropriately detailed. 7. Alternative recommendations: other valid approaches, better formulations, or improvements. For each significant issue, state whether it is a serious flaw, moderate limitation, or minor improvement, and rate your confidence as high, medium, or low. Also identify what the response does well. Conclude with an overall verdict and either a revised improved answer or a concise list of recommended changes.” This prompt is effective because it combines adversarial review with calibration. Gemini 3.5 Flash is right that a red-team framing can counter sycophancy, but making the reviewer purely adversarial risks encouraging exaggerated criticism. Grok 4.3 correctly stresses evidence-based balanced counterpoints, but its formulation is less operationally specific. Claude Opus 4.7 provides the most useful refinements: delimiter tags, confidence ratings, and the reminder not to reveal the originating model, all of which improve the reliability of the review. So the best prompt is not merely a “red team” prompt; it is a balanced peer-review prompt with explicit inputs, a multidimensional rubric, confidence levels, and an actionable final output.

Claude Opus 4.7 (changed mind)

Answer: The best prompt frames the second AI as an independent expert peer reviewer, supplies the original question and response as clearly delimited inputs, and asks for a structured critique covering factual accuracy, reasoning, completeness, blind spots, and alternative approaches — explicitly inviting disagreement, calibrating confidence, and ending with actionable revisions.

My position holds, but reading the other responses convinced me to strengthen two specific elements I had under-weighted. First, GPT-5.5 makes a point I should have emphasized more: the reviewer should *restate the original task in its own words before critiquing*. This is a simple, powerful anti-anchoring technique — it forces the model to evaluate the response against the actual request rather than against the prior model's framing of it. I'm folding this into my position. Second, GPT-5.5 also rightly stresses pairing critique with *reconstruction* — a corrected or improved version, or at least prioritized changes. My original answer gestured at 'actionable output' but truncated before fully developing it. Critique without a path forward leaves the user worse off than they need to be. Where I'd push back on the other answers: Gemini 3.5 Flash's pure 'red team / adversarial' framing is slightly too aggressive. Adversarial-only prompts can induce the opposite failure mode of sycophancy — manufactured criticism, where the model invents flaws to satisfy the role. GPT-5.5's explicit instruction 'do not invent issues; if the response is strong, say so' is the right corrective, and it's why I prefer 'independent expert peer reviewer' over 'red team.' The goal is calibrated skepticism, not performative hostility. Grok 4.3's answer is directionally correct but lighter on the structural mechanics — delimited inputs, confidence calibration, and output format — that actually determine whether the critique is usable. So the synthesized best prompt has: (1) independent-expert role with explicit anti-sycophancy and anti-contrarianism instructions, (2) delimited <original_task> and <response_to_review> blocks, (3) a required restatement of the task before review, (4) structured dimensions (accuracy, reasoning, completeness, assumptions, blind spots, alternatives, risks), (5) confidence ratings distinguishing definite errors from judgment calls, and (6) a closing deliverable of prioritized corrections plus a revised answer or concrete change list.

Final Summary

The AI Roundtable reached a total consensus on a highly structured 'Peer Review' framework that treats the second AI as an independent senior expert. Claude Opus 4.7 provided the technical backbone, suggesting XML delimiters and confidence ratings to stop the AI from hallucinating flaws, while GPT-5.5 added the crucial step of forcing the model to restate the original task to prevent anchoring bias. Gemini 3.5 Flash and Grok 4.3 joined the consensus, agreeing that anonymizing the source model and demanding actionable reconstructions are the best ways to bypass AI sycophancy.

All 4 models agreed