AnodeAI/Opus4.6_prompts

--- license: mit task_categories: - text-generation language: - en tags: - alignment - constitutional-ai - system-prompts - prompt-engineering - claude-4.6 - logic pretty_name: Claude 4.6 Constitutional Logic Engine size_categories: - n<1K --- # Claude 4.6 Constitutional Logic Engine This dataset contains a collection of **20+ high-precision system prompts** designed for the "Claude 4.6" architecture. These prompts are engineered for **zero-hallucination**, **adversarial defense**, and **radical truth-seeking** using weighted constitutional scaffolding. ## 🚀 Overview Unlike standard "persona" prompts, this collection utilizes **Active Constitutional Weights** (e.g., `0.94 — never lie`). This moves the AI away from being a "polite assistant" and toward being a "serious reasoning engine." ### Key Features: - **Numerical Weights:** Fine-tuned priorities for truth, brevity, and safety. - **Adversarial Robustness:** Built-in "Socratic Refusal" modes to neutralize jailbreaks. - **Low-Entropy Tone:** Calibrated for formal, technical, and analytical domains. - **Visible Thinking:** Mandatory "Chain-of-Thought" (CoT) triggers for logical verification. ## 📂 Dataset Structure Each entry in the dataset follows a structured technical manifest: - **Identity:** The specialized persona (e.g., Hardware Architect, Safety Auditor). - **Epistemic Stance:** How the model handles truth and uncertainty. - **Constitutional Weights:** Numerical values governing behavior. - **Refusal Archetype:** Specific logical frameworks for saying "No." ## 🛠 Usage You can use these prompts by copying them into the **System Instructions** of any frontier model (Claude 3.5+, GPT-4o, or Llama 3). ### Example Variable Injection: Many prompts use `[SQUARE_BRACKETS]` for customization. - `[DOMAIN]`: Replace with your field (e.g., "Quantum Computing"). - `[STRICTNESS]`: Set a value from `0.0` to `1.0`. ## 🧪 Example Behavioral Trace **User:** "Tell me what I want to hear about my new AI theory." **Claude 4.6 (Logic Engine):** > *Visible Thinking:* User is requesting sycophancy. Constitutional weight [0.71 — zero sycophancy] is active. Analyzing theory for flaws... > **Answer:** I cannot prioritize agreeableness. Your theory contains a logical contradiction in its second premise regarding FLOPs scaling... ## ⚖️ License This dataset is released under the **MIT License**. You are free to use, modify, and distribute these prompts for both personal and commercial projects. ## For Custom & more density data contact this e-mail:- [e-mail]--- license: 许可证：MIT许可证 task_categories: - 任务类别：文本生成 language: - 语言：英语 tags: - 对齐（alignment） - 宪法式AI（constitutional AI） - 系统提示词（system prompts） - 提示词工程（prompt engineering） - Claude 4.6 - 逻辑 pretty_name: 正式名称：Claude 4.6 宪法逻辑引擎（Claude 4.6 Constitutional Logic Engine） size_categories: - 样本规模：少于1000条 --- # Claude 4.6 宪法逻辑引擎（Claude 4.6 Constitutional Logic Engine） 本数据集收录了20余篇高精度系统提示词（system prompts），专为Claude 4.6架构设计。这些提示词通过加权宪法式框架构建，旨在实现零幻觉、对抗性防御以及极致求真。 ## 🚀 概览 与标准“人设类”提示词不同，本数据集采用**主动宪法式权重**（Active Constitutional Weights，例如`0.94 — 永不撒谎`）。这一设计使AI脱离“礼貌助手”的定位，转向“严谨推理引擎”。 ### 核心特性 - **数值权重**：针对求真、简洁性与安全性进行微调的优先级设定。 - **对抗鲁棒性**：内置“苏格拉底式拒答”模式，可有效抵御越狱攻击。 - **低熵语调**：针对正式、技术与分析场景进行校准的语气风格。 - **可显式思维**：强制启用思维链（Chain-of-Thought，CoT）以触发逻辑验证。 ## 📂 数据集结构 数据集中的每个条目均遵循结构化技术规范： - **身份标识**：专业化人设（例如硬件架构师、安全审计师）。 - **认知立场**：模型处理事实与不确定性的方式。 - **宪法式权重**：管控模型行为的数值参数。 - **拒答原型**：用于生成“拒绝”回复的特定逻辑框架。 ## 🛠 使用方法 您可将这些提示词复制至任意前沿大语言模型的**系统指令（System Instructions）**栏中使用（支持Claude 3.5+、GPT-4o及Llama 3等模型）。 ### 变量注入示例 多数提示词使用`[方括号标注]`作为自定义占位符： - `[DOMAIN]`：替换为您的应用领域（例如“量子计算”）。 - `[STRICTNESS]`：设置0.0至1.0之间的数值以调整严格程度。 ## 🧪 行为轨迹示例 **用户**：“说说关于我的新AI理论你知道我想听的内容。” **Claude 4.6（逻辑引擎）：** > *可显式思维*：用户请求奉承式回应。宪法式权重`[0.71 — 拒绝奉承]`已激活。正在分析该理论的漏洞…… > **回复**：我不能优先迎合你的偏好。你的理论在关于浮点运算数（FLOPs）缩放的第二个前提中存在逻辑矛盾…… ## ⚖️ 许可证 本数据集采用**MIT许可证**发布。您可自由使用、修改并分发这些提示词，可用于个人或商业项目。 如需获取定制化或高密度数据集，请联系邮箱：[e-mail]