Objective Critical AI

Problem It Addresses

Large language models often follow the user's framing, treat unstated assumptions as facts, present correlation as causation, or fill gaps with plausible but unsupported details. In engineering, research, and compliance work, that can corrupt downstream decisions. This skill constrains the answer path into a checkable sequence: inspect the user's assumptions, answer with evidence, decompose the assistant's own claims, and then critique the reasoning.

How It Works and When to Use It

• Assumption check: Extract the core claims in the input and test for missing constraints, causal confusion, or incomplete information. If an assumption is wrong, ask for confirmation or correction before answering.
• Evidence-backed answer: Answer only from validated assumptions and require evidence for each claim; when unsure, say uncertain instead of inventing details.
• Claim decomposition: Split the conclusion into independently verifiable facts and mark each as passed, failed, or questionable.
• Meta-critique: Check for logic jumps, missing counter-evidence, and overgeneralization, then output a revised answer.

It fits design reviews, paper discussion, incident analysis, and data-claim verification where reducing sycophancy matters. It does not replace domain expertise and cannot prove unsupported conclusions; when key facts are missing, gather evidence before asking for a definitive answer.