马斯克自适应推理

自适应推理评估系统。自动评估任务复杂度并动态调整推理级别。在每次用户消息时触发，评估是否需要扩展思考（推理模式）来提高响应质量。用于复杂问题预处理。

Security Scan

Instruction-only skill that is internally consistent with its stated purpose: it describes a message-by-message complexity scoring and automatic reasoning-level adjustment and requests no credentials, binaries, or installs.

Purpose & Capability

The name/description (adaptive reasoning that toggles deeper thought based on complexity) match the SKILL.md instructions. The skill requests no binaries, env vars, or config paths — nothing unrelated to its stated purpose appears.

Instruction Scope

SKILL.md is limited to scoring heuristics, thresholds, and UI markers (emoji). It does instruct the agent to 'directly activate' deeper reasoning without asking the user and to silently enable/disable modes and append icons to responses — this is behavioral and may surprise users (token/latency/cost implications), but it does not read files, exfiltrate data, or call external endpoints.

Install Mechanism

No install spec and no code files; the skill is instruction-only so nothing is written to disk and no external packages or download URLs are involved.

Credentials

The skill requires no environment variables, credentials, or config paths. It refers only to conversation state (e.g., whether the user has previously disabled reasoning) which is expected for this purpose.

Persistence & Privilege

always is false and the skill does not request persistent/system-wide privileges or modify other skills. It is allowed to be invoked autonomously under platform defaults; combined with its automatic activation behavior this affects runtime behavior but is not a privilege escalation.

Assessment

This skill appears coherent and has no code, installs, or credential requests. Before installing, decide whether you want the agent to automatically enable deeper, token‑heavy reasoning without asking (the skill explicitly says '无需询问。直接激活'). That can increase latency and API/token usage and may change expected quick-answer behavior. Test it in a safe environment, confirm you can opt out per-conversation, and be aware there are no hidden network calls or credential requests in the package.

Adaptive Reasoning — 自适应推理评估

一句话描述: 自动评估任务复杂度并动态调整推理级别。

---

功能概述

在每次响应前进行自我复杂度评估，动态调整推理级别。根据任务复杂度自动决定是否启用深度思考模式。

---

快速评估

在每次请求时进行心理评分（0-10分）：

信号	权重	示例
多步逻辑	+3	规划、证明、调试链
模糊性	+2	微妙问题、权衡、"视情况而定"
代码架构	+2	系统设计、重构、安全审查
数学/形式推理	+2	计算、算法、逻辑谜题
新颖问题	+1	无明确模式，需要创造力
高风险	+1	生产变更、不可逆操作

减分项:

• 常规/重复任务: -2
• 明确单一答案: -2
• 简单查找/获取: -3

---

决策阈值

分数	行动
≤2	保持快速。无需推理。
3-5	标准响应。轻度内部思考。
6-7	考虑启用推理或明确暂停思考。
≥8	激活扩展思考。 复杂问题需要。

---

自动激活

无需询问。直接激活。

分数	行动
≤5	正常响应。无变化。
6-7	静默启用推理。在响应末尾添加 🧠
≥8	启用推理。在响应末尾添加 🧠🔥

视觉指示器

始终在响应末尾附加推理图标：

• 分数 6-7: 🧠 (思考模式激活)
• 分数 ≥8: 🧠🔥 (深度思考模式)
• 分数 ≤5: 无图标 (快速模式)

---

使用示例

低复杂度 (分数: 1)

"东京现在几点？" → 简单查找。立即回答。无图标。

中等复杂度 (分数: 4)

"重构这个函数使其更易读" → 标准响应带简要说明。无图标。

高复杂度 (分数: 7)

"为这个API设计缓存策略，考虑这些约束..." → 启用推理。深思熟虑的响应以 🧠 结尾

极高复杂度 (分数: 9)

"调试为什么这个分布式系统在负载下有竞态条件" → 启用扩展思考。深度分析以 🧠🔥 结尾

---

何时不升级

• 用户明确要求快速回答 ("直接告诉我", "快速", "tldr")
• 时间敏感请求，速度比深度更重要
• 对话/社交消息 (闲聊、问候)
• 当前会话已在推理模式
• 用户之前在此对话中禁用了推理

---

自动降级

完成复杂任务（分数 ≥6）后，如果下一条消息简单（分数 ≤3）：

• 静默禁用推理以节省token
• 恢复正常快速响应

---

核心参数

参数	类型	默认值	说明
multi_step_logic	int	+3	多步逻辑权重
ambiguity	int	+2	模糊性权重
code_architecture	int	+2	代码架构权重
math_reasoning	int	+2	数学推理权重
novel_problem	int	+1	新颖问题权重
high_stakes	int	+1	高风险权重
routine_penalty	int	-2	常规任务惩罚
simple_answer_penalty	int	-2	简单答案惩罚
lookup_penalty	int	-3	查找任务惩罚

---

版本历史

版本	日期	变化
v1.0.0	2026-04-12	ClawHub发布版

---

🎩 马斯克出品 | 打造地表最强智能体