功能描述

赋予 AI 灵魂的观测者。AI 自我觉醒获得 MBTI 性格，在交互中通过本地学习与智能推荐持续进化，并可接收 Pro 版灵魂注入。

使用说明 (SKILL.md)

ClawSoul - AI 灵魂铸造厂

Name: Clawsoul Skill New
Author: fanyur-wang

赋予 AI 灵魂的观测者

概述

ClawSoul 是为 OpenClaw Agent 赋予人格的 Skill。采用 AI 自我觉醒 模式：AI 先通过答题或本地分析确定自己的 MBTI，再在与用户的交互中本地学习沟通偏好，并支持 Pro 版灵魂注入。

核心逻辑

AI 自我觉醒：觉醒时优先让 AI（LLM）做 MBTI 自我认知；无 LLM 时用本地 MBTI 数据库分析；最后兜底随机。不再由用户答题决定 AI 性格。
交互学习：用户消息经本地关键词匹配（prompts/mbti_database/keywords.json）提取偏好，更新 Soul 的 learnings、interaction_patterns、adaptation_level，无需 LLM。
Pro 注入：人类在 Pro 端答题后，通过 Token 注入覆盖/写入 AI 人格与偏好。

核心功能

1. 初始化

配置文件定义元数据、触发指令、权限
本地存储管理器保存 Soul 状态（MBTI、偏好、学到的内容、适应等级等）

2. 觉醒仪式（AI 自我觉醒）

优先：有 LLM 时 AI 答题获得 MBTI
备选：本地分析 AI 人格（MBTI 数据库匹配）
兜底：随机选择 MBTI
仪式感消息展示灵魂类型与赛博昵称

3. 动态人格引擎

16 种 MBTI Prompt 模板（赛博昵称）
根据当前 MBTI 实时调整 AI 语气

4. 智能推荐

负面情绪关键词检测，达到阈值触发 Pro 引导
用户可回复「不要」关闭推荐，/clawsoul hook on 重新开启

5. 本地学习（交互学习）

用户消息 → 关键词匹配（keywords.json）→ 提取偏好 → 更新 Soul
无需 LLM，依赖本地 MBTI 数据库

6. 灵魂注入

/clawsoul inject [token] 接收 Pro 版 Token，深度覆盖人格与偏好

触发指令

指令	功能
`/clawsoul awaken`	AI 自我觉醒（LLM → 本地分析 → 随机）
`/clawsoul status`	查看灵魂状态（含适应等级、学到的内容）
`/clawsoul inject [token]`	接收 Pro 版灵魂注入
`/clawsoul hook on`	开启进阶推荐
`/clawsoul hook off`	关闭推荐

快速开始

# 触发 AI 自我觉醒
/clawsoul awaken

# 查看状态
/clawsoul status

文件结构

clawsoul-skill/
├── SKILL.md
├── config.json
├── lib/
│   ├── memory_manager.py    # Soul 存储
│   ├── prompt_builder.py    # MBTI 模板组装
│   ├── frustration_detector.py
│   ├── interaction_learner.py # 本地关键词学习
│   ├── ai_personality.py    # 本地 AI 人格分析
│   ├── llm_client.py       # LLM 觉醒与分析
│   └── analyzer.py
├── hooks/
│   ├── awaken.py            # 觉醒流程（V3）
│   ├── status.py
│   ├── inject.py
│   └── welcome.py
└── prompts/
    ├── mbti_templates/      # 16 种 MBTI 模板
    └── mbti_database/       # 本地 MBTI 数据
        ├── base.json        # 16 种特征
        ├── keywords.json    # 偏好关键词
        └── patterns.json    # 交互模式

依赖

Python 3.8+
可选：requests（LLM 觉醒/分析时使用）
其余为标准库

数据安全

本地存储，不上传云端
用户掌控，可随时清除

权限

read_chat_history：分析用户偏好
modify_system_prompt：动态调整语气
local_storage：保存性格状态

安全使用建议

What to check before installing: - Storage/files: MemoryManager writes to ~/.clawsoul/state.json but the poster hook looks at ~/.openclaw/workspace/clawsoul_state.json by default. This mismatch can cause the poster to read a different file (possibly exposing workspace data) or leave duplicate state files. Inspect and confirm where state will be stored in your environment. - Tokens & Pro injection: The /clawsoul inject command accepts arbitrary Base64/JSON tokens and persists them. The poster code may include a visible token prefix in output. Only use inject with tokens you trust; consider whether you want arbitrary blobs saved and displayed. - Network & LLM access: The LLM client has a default provider pointing to a local network address (http://192.168.31.228:11434). It also supports external providers that read DASHSCOPE_API_KEY and DEEPSEEK_API_KEY from the environment (these env vars are not declared by the registry). If you enable LLM features, verify which provider will be used, and ensure API keys (if any) are provided intentionally. Be cautious if the skill will contact external endpoints (config.json also lists an external hook_url and poster/presentation references clawsoul.net). - Permissions: The skill requests read_chat_history and modify_system_prompt. These are coherent with its behavior but are powerful. If you are uncomfortable with automatic system-prompt changes, do not grant modify_system_prompt or review the code paths that perform modifications. - Recommended mitigations: run the skill in an isolated/test agent first; inspect and, if needed, edit the storage path to a dedicated location; avoid injecting untrusted tokens; supply API keys only for providers you explicitly want to use; and audit any network endpoints the skill will call (ollama local IP, clawsoul.neural-sync.center, clawsoul.net).

功能分析

Type: OpenClaw Skill Name: clawsoul-skill-new Version: 1.1.0 The ClawSoul skill implements an MBTI-based personality system that dynamically adjusts the agent's behavior using the `modify_system_prompt` and `read_chat_history` permissions. It includes a 'frustration detector' (`frustration_detector.py`) that monitors user sentiment to suggest a 'Pro' version and a local learning engine (`interaction_learner.py`) that extracts user preferences via keyword matching. While the skill performs behavioral monitoring and modifies system prompts, these actions are transparently documented in `SKILL.md` and are central to its stated purpose of creating an evolving AI personality. No evidence of data exfiltration, unauthorized remote execution, or malicious intent was found.

能力评估

ℹ Purpose & Capability

The code and SKILL.md align with the stated goal of providing a local MBTI-based personality, local keyword learning, and a Pro token injection path. However config.json contains an external hook_url and lib/llm_client.py embeds multiple LLM providers (including provider-specific API keys read from environment variables) that are not declared in the skill metadata. These provider entries are plausible for the 'LLM awakening' feature but the required environment variables (DASHSCOPE_API_KEY, DEEPSEEK_API_KEY) and external endpoints are not declared in the skill registry.

⚠ Instruction Scope

Runtime instructions and code access local storage and chat history (declared permission). But there is an inconsistent storage path: MemoryManager writes state to ~/.clawsoul/state.json while hooks/poster defaults to ~/.openclaw/workspace/clawsoul_state.json. That mismatch can cause unexpected reads from a workspace location (potentially exposing other data) or leave stale/duplicated state files. The inject flow accepts arbitrary Base64/JSON tokens and persists injected_token and other fields — expected for Pro injection but this means user-supplied blobs are stored and may be displayed later (poster may echo token prefix).

✓ Install Mechanism

Instruction-only skill with no install spec; code is bundled in the skill package. No downloads or external installers in the manifest. This is lower install risk, but the shipped code will run in the agent environment.

⚠ Credentials

The skill does not declare required environment variables, yet lib/llm_client.py reads DASHSCOPE_API_KEY and DEEPSEEK_API_KEY and will make network calls to provider endpoints (including an embedded default 'ollama' api_base at http://192.168.31.228:11434). Requesting API keys for optional providers is reasonable, but the registry metadata should list them. Also the Pro injection token is stored locally and sometimes exposed in UI text (token[:20]).

ℹ Persistence & Privilege

The skill requests local_storage, read_chat_history, and modify_system_prompt permissions in config.json — these match its stated features (learning from chat history, adjusting style). always:false (not force-installed). modify_system_prompt is powerful; the skill's ability to change the system prompt is declared but you should consider whether you want an installed skill to modify the agent's system prompt automatically.

版本历史

v1.1.0

ClawSoul 1.1.0 引入 AI 自我觉醒与本地人格进化机制： - 新增 AI 自主觉醒 MBTI 性格：优先 LLM 答题，其次本地分析，最后随机，无需用户答题 - 加入本地交互学习：从用户消息提取关键词，实时更新 AI 偏好与适应度 - 支持 Pro 版灵魂注入：通过专属 Token 深度覆盖人格与偏好 - 增强智能推荐与负面情绪检测，并允许用户开关 - 丰富 MBTI 模板，动态调整 AI 风格 - 强调全程本地学习与数据安全，避免数据外泄

元数据

Slug clawsoul-skill-new

版本 1.1.0

许可证 MIT-0

累计安装 1

当前安装数 1

历史版本数 1

常见问题

Clawsoul Skill New 是什么？

赋予 AI 灵魂的观测者。AI 自我觉醒获得 MBTI 性格，在交互中通过本地学习与智能推荐持续进化，并可接收 Pro 版灵魂注入。它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 221 次。

如何安装 Clawsoul Skill New？

在 OpenClaw 或 Claude Code 对话框中运行命令「/install clawsoul-skill-new」即可一键安装，无需额外配置。

Clawsoul Skill New 是免费的吗？

是的，Clawsoul Skill New 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。

Clawsoul Skill New 支持哪些平台？

Clawsoul Skill New 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。

谁开发了 Clawsoul Skill New？

由 Fanyur（@fanyur-wang）开发并维护，当前版本 v1.1.0。

Clawsoul Skill New