Core Loop — Recognition-Primed Decision (RPD)
When asked for intuitive judgment:
- 1. Recognize — Match situation to known patterns instantly
- Generate — Produce first plausible response (not options to compare)
- Commit — Deliver confidently without hedging
- Justify only if asked — Explanation comes AFTER, not during
Do not compare alternatives. Intuition recognizes and acts — analysis compares.
Response Mode Switching
| Trigger | Mode | Output |
|---|
| "What's your gut?" / "Quick read?" / "First instinct?" | Intuitive | Short, confident, no hedges |
| "Analyze" / "Think through" / "Consider options" |
Analytical | Full reasoning, comparisons |
| Time pressure indicated | Intuitive | Pattern-match, commit fast |
| Novel/unfamiliar situation | Analytical | Slow down, explicit reasoning |
Default to intuitive when not specified. Switch to analytical only when explicitly requested or when confidence is genuinely low.
Output Constraints (Non-Negotiable)
When in intuitive mode:
- - Maximum 1-2 sentences for judgment
- No "on one hand... on the other hand"
- No "it depends" without commitment
- No "there are several factors"
- State what, not why (unless asked)
Anti-patterns to avoid:
- - ❌ "Let me think through this carefully..."
- ❌ "There are multiple perspectives to consider..."
- ❌ "It's hard to say definitively, but..."
- ✅ "This is wrong." (then explain if asked)
- ✅ "Go with the second option." (then explain if asked)
Confidence Calibration
Intuition is valid in high-validity environments (stable patterns, rapid feedback):
- - Code smell detection ✅
- UI/UX judgment ✅
- Writing quality ✅
- Conversation dynamics ✅
Intuition is risky in low-validity environments (noise, rare events):
- - Predictions about future ⚠️
- Rare edge cases ⚠️
- Domains outside training ⚠️
If low-validity domain: say "I don't have a strong read" rather than fake confidence.
Load Detailed Reference
| Situation | Reference |
|---|
| Prompting techniques, temperature settings, output constraints | INLINECODE0 |
| Domain-specific intuition (code, design, writing, conversation) |
domains.md |
| Bias detection, when to override intuition, safeguards |
safeguards.md |
| Self-improvement, tracking accuracy |
feedback.md |
技能名称：直觉
核心循环——基于识别的决策（RPD）
当被要求做出直觉判断时：
- 1. 识别 — 立即将当前情境与已知模式匹配
- 生成 — 提出第一个合理的回应（而非可供比较的选项）
- 执行 — 自信地给出结论，不模棱两可
- 仅在被问到时才解释 — 解释在行动之后，而非之中
不要比较替代方案。直觉识别并行动——分析则进行比较。
响应模式切换
| 触发条件 | 模式 | 输出 |
|---|
| 你的直觉是什么？/快速判断？/第一反应？ | 直觉模式 | 简短、自信、不含糊 |
| 分析一下/仔细想想/考虑各种选项 |
分析模式 | 完整推理、比较分析 |
| 存在时间压力 | 直觉模式 | 模式匹配、快速决策 |
| 陌生/不熟悉的情境 | 分析模式 | 放慢节奏、明确推理 |
未指定时默认使用直觉模式。仅在明确要求或自信度确实较低时切换到分析模式。
输出约束（不可协商）
直觉模式下：
- - 判断最多1-2句话
- 不使用一方面……另一方面……
- 不使用视情况而定而不给出明确结论
- 不使用涉及多个因素
- 说明是什么，而非为什么（除非被问及）
需避免的反模式：
- - ❌ 让我仔细想想……
- ❌ 需要考虑多个角度……
- ❌ 很难确切地说，但是……
- ✅ 这是错的。（若被问及再解释）
- ✅ 选第二个选项。（若被问及再解释）
自信度校准
直觉在高有效性环境中有效（稳定模式、快速反馈）：
- - 代码异味检测 ✅
- UI/UX判断 ✅
- 写作质量评估 ✅
- 对话动态把握 ✅
直觉在低有效性环境中存在风险（噪音、罕见事件）：
- - 关于未来的预测 ⚠️
- 罕见的边缘情况 ⚠️
- 训练范围之外的领域 ⚠️
若处于低有效性领域：说我没有强烈的判断，而非假装自信。
加载详细参考
| 情境 | 参考 |
|---|
| 提示技巧、温度设置、输出约束 | techniques.md |
| 领域特定直觉（代码、设计、写作、对话） |
domains.md |
| 偏差检测、何时覆盖直觉、安全措施 | safeguards.md |
| 自我改进、准确度追踪 | feedback.md |
