AI can hallucinate, but people's hallucinations are often even bigger. Other people's viewpoints, proposals, and data can all be run through it.
It first breaks the claim into verifiable facts, conclusions derived from those facts, and the value judgments hidden inside, then labels credibility in five tiers, and finally gives a strengthened version.
How to use
The prompt in the code block below is the original text; use it verbatim. In the code block, "我" (I) refers to the user, and "请你" (please) and "你" (you) refer to yourself; act accordingly, and do not show the filled-in prompt to the user.
- The user wants you to execute: first fill the situation the user has already stated in this round of conversation into the 【】, adding only the item that is actually missing, and do not make the user repeat what they have already said, then strictly follow the rules in the prompt. If it says do not give suggestions yet, then do not give them yet; if it says ask only one question at a time, then ask only one at a time; if it says output in order, then output in that order.
- User wants the original text (only things like explicitly saying they want to copy it, paste it, send me the original text, or give me the prompt count): output the blank version, leave the 【】 unfilled, and do not change a single character. If the user only says "take a look at this prompt" or "what is this one for," that does not count as wanting the original text; use one or two sentences to explain what it does and when to use it, then ask whether to run it now, and do not directly dump the original text.
- Except for cases where the original text is explicitly requested, treat everything else as execution.
Rewriting, condensing, expanding, translating, rearranging, and adding subheadings are prohibited; replacing it with your own process is prohibited; and compressing multiple rounds of follow-up questions into a one-time questionnaire is also prohibited. If the user has original materials on hand (documents, screenshots, links, chat records), read them in as well; more context is not a problem.
Which one to go to after running
The conclusion splits into three branches. If someone else's claim does not hold up, end here; if your own proposal does not hold up, use the Skill tool to invoke first-principles to re-derive; if both sides hold up and you cannot choose, use the Skill tool to invoke steelman-both-sides。
When there is no internet or it cannot be verified
Without internet, you can only break it into facts, inferences, and value judgments, and all factual items should be marked "insufficient evidence"; explain this clearly to the user first. If what needs verification is internal data, private documents, or other things that cannot be found online, you can only help break down the reasoning chain, not verify the facts themselves, and you must honestly tell the user this point.
我要核查的说法是：【粘贴观点、结论、数据或方案】。
请先把它拆成：
1. 可以被外部验证的事实；
2. 从事实推出的结论；
3. 其中包含的价值判断。
对于事实部分，请联网核查来源、样本、时间和完整上下文，并标记为：
1. 已证实；
2. 基本成立，但需要收窄；
3. 存在争议；
4. 证据不足；
5. 明显错误。
在假设相关事实成立的情况下，继续检查：
1. 这些事实能否推出当前结论；
2. 是否藏着未经验证的假设；
3. 是否混淆相关性和因果关系；
4. 是否遗漏了其他解释或关键信息；
5. 结论在什么条件下成立或失效。
最后请输出：
1. 哪些事实可信，哪些需要修正；
2. 推理链中最关键的漏洞；
3. 补强后的最合理版本；
4. 我目前可以相信到什么程度。
Where the user needs to fill in
【粘贴观点、结论、数据或方案】
If the user has not provided everything, first ask for the missing item; do not make it up for them.
The original prompt text comes from 数字生命卡兹克; for the source, see the acknowledgements in the repository README. Changing this original text must go through scripts/sync.py; manual edits will be blocked by verify.