Anti-Hallucination Pro

What problem it addresses

Large models can turn training memory, user assumptions, or plausible-sounding reasoning into confident answers. This is especially risky when handling verifiable facts, data, regulations, causal claims, time-sensitive information, niche knowledge, API/library choices, environment dependencies, or feasibility assessments. A statement such as “it probably needs installation” or “it likely does not support this” can become an incorrect technical decision. This skill treats those situations as risk cases and requires the agent to check premises, verify sources, and only then conclude.

How the skill works

• Trigger and exemption: It activates for verifiable facts, explicit fact-checking requests, low-frequency niche knowledge, and tool selection/solution decisions. Pure creative writing, translation, formatting, and open-ended discussion are not triggered by default, but real-world citations and API/library choices still follow the rules.
• Verify before answering: Factual information should be checked with tools rather than relying on training data. Time-sensitive information must be retrieved in real time. Multi-turn citations should be traced back to the original source. Before tool or solution decisions, key assumptions should be validated through actions such as pip install, curl, or file inspection.
• Label reasoning levels: Separate facts, inferences, and guesses. Specific numbers, dates, names, papers, and legal clauses require sources; otherwise use directional language. Confidence is graded using prior probability and evidence exclusivity into levels such as certain, very likely, possible, uncertain, and unknown.
• Two-pass self-review: Pass 1 drafts the answer with source and confidence labels. Pass 2 reviews it from a skeptical perspective. Higher-depth review is used for decisions, high-stakes domains, strict fact-checking requests, or skill review scenarios.
• Verification routing: Choose among source tracing, self-checking, chained validation, partitioning, auditing, or step-back reasoning. When search results are insufficient, the answer should state what is missing and suggest better query directions. When sources conflict, present the discrepancy instead of choosing a side.

Boundaries and cautions

This is a prompt-level behavioral constraint, not an engine-level hard block, so its effect depends on the agent’s compliance. It is useful for fact checking, technical selection, risk-sensitive answers, and review workflows. It is not intended for fully open creative writing without factual citations. If the user asks for answers from internal understanding only, the output can be provided but should be marked as not real-time verified, with key data still carrying confidence labels.