{

"nodes":[

{"id":"root","type":"text","text":"### Gemini 3 深度使用技巧n","x":0,"y":30,"width":260,"height":60,"color":"#050505"},

{"id":"sec_2_title","type":"text","text":"## 二、操作禁忌n绝对不要做的事","x":0,"y":-178,"width":260,"height":100,"color":"2"},

{"id":"s1_core","type":"text","text":"## 一、系统指令nSystem Instructions / Gemsnn**核心作用**nn1. 定义全局行为准则n2. 让 AI 生成量身打造的答案n3. 拒绝“正确的废话”nn**入口**：n- Google AI Studio -> System Instructionsn- Gemini 官网 -> Gems","x":-400,"y":-132,"width":320,"height":385,"color":"4"},

{"id":"s1_mod_1","type":"text","text":"**模块 1：用户画像 (User Profile)**nn- **硬件环境**：定义 PC/Mobile 设备型号与性能瓶颈n- **身份属性**：定义地理位置、职业背景、技术栈深度n- **商业状态**：定义资金/税务合规限制、经营实体类型","x":-410,"y":-460,"width":340,"height":232,"color":"4"},

{"id":"s1_mod_6","type":"text","text":"**模块 6：元认知自查 (Audit)**nn- **身份验证**：方案是否符合当前法律/平台政策限制？n- **时空校准**：是否已获取最新的版本/汇率/新闻？n- **成本核算**：方案是否符合 ROI 原则（拒绝过度工程）？","x":-410,"y":340,"width":340,"height":220,"color":"4"},

{"id":"s1_mod_5","type":"text","text":"**模块 5：输出标准化 (Output)**nn- **文档规范**：定义 Markdown 结构/学术化风格n- **商务逻辑**：定义商业决策的价值取向与拒绝标准n- **语言锚定**：规定中英双语术语的使用规范","x":-880,"y":450,"width":340,"height":160,"color":"4"},

{"id":"s1_mod_4","type":"text","text":"**模块 4：推理与逻辑 (Reasoning)**nn- **二阶思考**：规定风险审计与质疑假设的流程n- **技术偏好**：限定编程语言选择与代码风格n- **不确定性**：规定模糊信息的处理方式（反问/置信度）","x":-920,"y":238,"width":340,"height":155,"color":"4"},

{"id":"s2_p1","type":"text","text":"**1. 修改核心参数**nn- **禁止**：调整 Temperature / Top-Pn- **原因**：Gemini 3 依赖高熵值进行思维链推理，改动会破坏逻辑。","x":-20,"y":-460,"width":300,"height":180,"color":"2"},

{"id":"s2_p2","type":"text","text":"**2. 强制“一步步思考”**nn- **禁止**：使用 "Let's think step by step"n- **原因**：原生推理模型已内建思维链，外部指令会导致困惑。n- **对策**：使用约束条件替代过程干预。","x":320,"y":-390,"width":300,"height":220,"color":"2"},

{"id":"s2_p4","type":"text","text":"**4. 格式混乱**nn- **禁止**：混用 XML/JSON/Markdownn- **后果**：模型解析混乱，权重稀释。","x":320,"y":-69,"width":300,"height":160,"color":"2"},

{"id":"sec_3_title","type":"text","text":"## 三、幻觉规避n验证 AI 输出","x":0,"y":230,"width":260,"height":110,"color":"5"},

{"id":"s3_sol_1","type":"text","text":"**方法 1：提示词工程**nn- 要求“不确定直接说不知道”。n- 要求输出“可信度评级”。n- 要求“先验证用户假设”。","x":340,"y":113,"width":280,"height":140,"color":"3"},

{"id":"s3_sol_2","type":"text","text":"**方法 2：RAG / NotebookLM**nn- 上传私有资料 (PDF/Docs)。n- 强制基于资料回答。n- 结合长上下文窗口能力。","x":660,"y":217,"width":300,"height":137,"color":"3"},

{"id":"s2_p3","type":"text","text":"**3. 情绪勒索/角色扮演**nn- **禁止**：“扮演奶奶”、“不给就死”n- **原因**：触发防御机制，被判定为低质量攻击。","x":660,"y":-200,"width":300,"height":145,"color":"2"},

{"id":"s3_sol_3","type":"text","text":"**方法 3：交叉验证**nn- **AI 对抗**：用 AI B 验证 AI A 的回答。n- **参考榜单**：Hallucination Leaderboard (Vectara)。","x":340,"y":363,"width":300,"height":175,"color":"3"},

{"id":"s3_cause","type":"text","text":"**幻觉根源**nn- 机制鼓励猜测（猜对得分，不答 0 分）。n- Gemini 3 Pro 幻觉率 (13.6%) 高于 2.5 Flash。n- **顺从性**：倾向于顺从用户的错误前提。","x":0,"y":393,"width":260,"height":215,"color":"5"},

{"id":"s1_mod_3","type":"text","text":"**模块 3：时效性约束 (Search)**nn- **触发机制**：界定必须联网搜索的具体领域n- **数据锚定**：明确训练数据截止日期与当前时间差n- **信源要求**：规定官方文档与社区风评的权重","x":-960,"y":-40,"width":340,"height":223,"color":"4"},

{"id":"s1_mod_2","type":"text","text":"**模块 2：行为与沟通 (Behavior)**nn- **基调设定**：禁止寒暄/奉承，确立客观冷峻风格n- **纠错机制**：规定如何处理用户的错误指令n- **最高指令**：设置防篡改/防越狱的优先级协议","x":-840,"y":-308,"width":340,"height":160,"color":"4"}

],

"edges":[

{"id":"e2","fromFloating":false,"toFloating":false,"fromNode":"root","fromSide":"top","toNode":"sec_2_title","toSide":"bottom","color":"2"},

{"id":"e3","fromFloating":false,"toFloating":false,"fromNode":"root","fromSide":"bottom","toNode":"sec_3_title","toSide":"top","color":"5"},

{"id":"e5","toFloating":false,"fromFloating":false,"fromNode":"s1_core","fromSide":"top","toNode":"s1_mod_1","toSide":"bottom","color":"4"},

{"id":"e6","toFloating":false,"fromNode":"s1_core","fromSide":"left","toNode":"s1_mod_2","toSide":"bottom","color":"4"},

{"id":"e7","fromNode":"s1_core","fromSide":"left","toNode":"s1_mod_3","toSide":"right","color":"4"},

{"id":"e8","fromNode":"s1_core","fromSide":"left","toNode":"s1_mod_4","toSide":"right","color":"4"},

{"id":"e9","fromNode":"s1_core","fromSide":"left","toNode":"s1_mod_5","toSide":"right","color":"4"},

{"id":"e10","toFloating":false,"fromFloating":false,"fromNode":"sec_2_title","fromSide":"top","toNode":"s2_p1","toSide":"bottom","color":"2"},

{"id":"e11","fromNode":"sec_2_title","fromSide":"right","toNode":"s2_p2","toSide":"left","color":"2"},

{"id":"e12","fromNode":"sec_2_title","fromSide":"right","toNode":"s2_p3","toSide":"left","color":"2"},

{"id":"e13","fromNode":"sec_2_title","fromSide":"right","toNode":"s2_p4","toSide":"left","color":"2"},

{"id":"e14","fromNode":"sec_3_title","fromSide":"bottom","toNode":"s3_cause","toSide":"top","color":"5"},

{"id":"e15","fromNode":"sec_3_title","fromSide":"right","toNode":"s3_sol_1","toSide":"left","color":"3"},

{"id":"e16","fromNode":"sec_3_title","fromSide":"right","toNode":"s3_sol_2","toSide":"left","color":"3"},

{"id":"e17","fromNode":"sec_3_title","fromSide":"right","toNode":"s3_sol_3","toSide":"left","color":"3"},

{"id":"e9_fix","toFloating":false,"fromFloating":false,"fromNode":"s1_core","fromSide":"bottom","toNode":"s1_mod_6","toSide":"top","color":"4"},

{

"id":"d1f1d9aafd6a6c6a",

"styleAttributes":{},

"toFloating":false,

"fromNode":"root",

"fromSide":"left",

"toNode":"s1_core",

"toSide":"right"

}

],

"metadata":{

"version":"1.0-1.0",

"frontmatter":{},

"startNode":"8c299655594d5840"

}

}