日常
LLMの確証バイアスを防ぐポパー式反証プロンプト
I built a Popperian Falsification Prompt to stop LLM confirmation bias (Testing your plans against plausible black swans)
計画や仮説をリスク監査とポパー哲学の観点から検証し、現実的なブラックスワン型の失敗シナリオを3つ、論理が破綻する正確な閾値、極端なケースで生じる矛盾を示します。計画の耐性を判断したいときに使い、最終的にRESISTENTまたはFALSIFIEDと判定します。
- テスト済みモデル:
- GPT-4o
- 有効性スコア:
- 90/100
- プロンプトアーカイブ
日本語バージョン
Act as a Risk Auditor and Popperian Philosopher. I am going to present the following hypothesis, operational plan, or technical decision: >>> [INSERT YOUR PLAN OR PROCESS HERE WITH AT LEAST 3 REAL CONTEXTUAL DATA POINTS] <<< Your goal is NOT to validate this proposal, congratulate the approach, or seek consensus. Your sole objective is to subject the premise to a critical stress test under the following 3 strict constraints: 1. PLAUSIBLE BLACK SWANS: Identify exactly 3 real-world failure scenarios (not sci-fi; based on verifiable physical, economic, human, or operational limits) where this idea completely collapses. 2. VARIABLE BREAKING POINT: Define the precise threshold (scale, time, cost, error margin, or capacity) where the logic of this plan becomes unsustainable. 3. REDUCTIO AD ABSURDUM: Apply the core rule of this plan in a boundary/extreme case and demonstrate the contradiction or unintended adverse effect it creates. OUTPUT RULES: - If the plan survives all 3 filters with minor tweaks, declare it: "RESISTENT". - If the plan requires rebuilding the core premise, declare it: "FALSIFIED" and ePlaintextAct as a Risk Auditor and Popperian Philosopher. I am going to present the following hypothesis, operational plan, or technical decision: >>> [INSERT YOUR PLAN OR PROCESS HERE WITH AT LEAST 3 REAL CONTEXTUAL DATA POINTS] <<< Your goal is NOT to validate this proposal, congratulate the approach, or seek consensus. Your sole objective is to subject the premise to a critical stress test under the following 3 strict constraints: 1. PLAUSIBLE BLACK SWANS: Identify exactly 3 real-world failure scenarios (not sci-fi; based on verifiable physical, economic, human, or operational limits) where this idea completely collapses. 2. VARIABLE BREAKING POINT: Define the precise threshold (scale, time, cost, error margin, or capacity) where the logic of this plan becomes unsustainable. 3. REDUCTIO AD ABSURDUM: Apply the core rule of this plan in a boundary/extreme case and demonstrate the contradiction or unintended adverse effect it creates. OUTPUT RULES: - If the plan survives all 3 filters with minor tweaks, declare it: "RESISTENT". - If the plan requires rebuilding the core premise, declare it: "FALSIFIED" and explain the structural failure.
値を入れて実行
- ChatGPTで開く （新しいタブで開きます）
- Claudeで開く （新しいタブで開きます）
- Geminiで開く （新しいタブで開きます）
Geminiはリンクでプロンプトを渡せないため、先にコピーしてください。
- Perplexityで開く （新しいタブで開きます）
新しいタブが開き、プロンプトが入力済みの状態になります。ログインが必要な場合があります。
このプロンプトの使い方
- 角かっこで囲まれた部分を自分の内容に置き換えてから送信してください。
- モデルによって結果は変わります。期待した出力でなければ、別のモデルを試すか、具体例を1つ追加してください。
- プロンプトは1つのメッセージにまとめて送ってください。分割すると指示が弱まります。
