doublecheck

The doublecheck skill is designed to improve the reliability of AI-generated output by introducing a structured, multi-layer verification pipeline. Instead of simply producing answers, it focuses on identifying and validating factual claims within responses. It breaks down generated content into verifiable statements, then evaluates each one against externally discoverable evidence, primarily through web-based sources. The core objective is to reduce misinformation risks, especially in contexts where AI outputs may include unsupported or uncertain assertions.

The skill operates through a combination of mechanisms, including claim extraction, source-based verification, and adversarial review for hallucination patterns. It supports two operating modes: an active mode that continuously applies verification logic throughout a conversation, and a one-shot mode for verifying specific inputs on demand. In active mode, responses are classified by claim density, and verification is applied inline using structured labels such as VERIFIED, PLAUSIBLE, or FABRICATION RISK. When higher risk is detected, the system escalates from lightweight inline checks to a more comprehensive report format. It may also include a "Heads up" warning when confidence is low and supporting evidence is not found.

This skill is primarily useful in environments where correctness and traceability are critical. It is intended for users working with research summaries, factual analysis, regulatory or compliance-related content, and any scenario where AI-generated information must be independently checkable. It is also valuable for developers, analysts, educators, and writers who need transparent validation of claims. By structuring verification alongside generated content, doublecheck helps users better understand which parts of an answer are supported, uncertain, or potentially unreliable.