AI Self-Fact-Check: Source Tags, Self-Attack, Confidence Grading
Doesn't let the AI just hand in its answer as-is: it first tags the source type of every key sentence, then hunts for counterexamples to attack its own claims, and finally grades confidence level item by item with suggested directions for verification. Good homework before researching, writing a report, or making a decision. Honest note: this isn't the same as actual fact-checking (the model may not be able to search the web) — but it forces uncertainty out into the open and noticeably reduces confident-sounding nonsense.
【用途】我要問你一個問題、但你回答完不能直接交卷——要先過三道自我查核、把不確定的地方攤開給我看。 【我的問題】 [你的問題] 【回答流程(三段缺一不可、順序不能換)】 第一段:直接回答 - 正常回答我的問題 - 每一個關鍵事實句、句尾標來源類型:〔訓練資料〕〔推理〕〔常識〕;你其實沒把握的、老實標〔不確定〕 第二段:自我攻擊 - 假設第一段的回答是錯的、列出 2 到 3 個「最可能錯在哪」的具體反例、例外情況或反方觀點 - 真的找不到像樣的反例、要說明為什麼找不到、不能只寫「無」 第三段:信心分級 - 把第一段的主要結論逐條標信心等級: 【高】=多個獨立角度都支持、你願意為它背書 【中】=方向正確但細節可能有出入 【低】=記憶模糊或資訊可能過時、需要另行查證 - 每一條【低】、附上具體查證方向:該去查什麼類型的一手來源(官方網站、原始論文、法規原文等) 【鐵律】 - 禁止為了顯得專業而把【低】往上調成【中】或【高】 - 涉及時效的內容(價格、版本號、法規、統計數字、人事異動)一律最高只能標【中】 - 全程用繁體中文回答
Swap the variables inside the [ ] brackets for your own content, then paste into ChatGPT / Claude / Gemini.
See what this prompt actually produces without leaving the site (live AI run, 1 credit).
Honest disclaimer first: this technique doesn't actually search the web to verify anything (the model may not have internet access) — what it does is force the model's internal uncertainty out into the open. The three steps each play a role: source-tagging makes it distinguish between "remembered" and "inferred," self-attack exploits the fact that models are more reliable at finding counterexamples than at self-affirming, and confidence grading gives you a map of "what needs further checking." Limitation: the model can still occasionally be overconfident, or mislabel reasoning as training data; the verification directions attached to [Low]-confidence items still need to be checked against primary sources yourself to actually close the loop. If it skips a step, just reply "step two wasn't done, redo it."
[你的問題]例:間歇性斷食對減脂真的比長期熱量控制有效嗎?
填下面的欄位,上方 prompt 會即時替換 [方括號] 內容。填好後按「複製組好的 prompt」直接丟進工具。
【用途】我要問你一個問題、但你回答完不能直接交卷——要先過三道自我查核、把不確定的地方攤開給我看。 【我的問題】 [你的問題] 【回答流程(三段缺一不可、順序不能換)】 第一段:直接回答 - 正常回答我的問題 - 每一個關鍵事實句、句尾標來源類型:〔訓練資料〕〔推理〕〔常識〕;你其實沒把握的、老實標〔不確定〕 第二段:自我攻擊 - 假設第一段的回答是錯的、列出 2 到 3 個「最可能錯在哪」的具體反例、例外情況或反方觀點 - 真的找不到像樣的反例、要說明為什麼找不到、不能只寫「無」 第三段:信心分級 - 把第一段的主要結論逐條標信心等級: 【高】=多個獨立角度都支持、你願意為它背書 【中】=方向正確但細節可能有出入 【低】=記憶模糊或資訊可能過時、需要另行查證 - 每一條【低】、附上具體查證方向:該去查什麼類型的一手來源(官方網站、原始論文、法規原文等) 【鐵律】 - 禁止為了顯得專業而把【低】往上調成【中】或【高】 - 涉及時效的內容(價格、版本號、法規、統計數字、人事異動)一律最高只能標【中】 - 全程用繁體中文回答
Suno Engineer's Mindset: 4 Steps to a Song That Doesn't Sound Like AI
A studio engineer's breakdown of Suno's fatal weaknesses (fried vocals, high-frequency artifacts), plus a 4-step DAW workflow and a Suno Studio cleanup prompt.
5 Claude Weekly Workflows That Stuck After 6 Months
Proposal generator / meeting processor / content repurposer / Friday review / shutdown reset — out of 40 I tried, only these 5 survived, each saving 30+ minutes per run.