persona-creator

数字人风格生成器。当用户说"生成我的数字人"、"分析我的说话风格"、"创建我的风格档案"， 或输入 /superme 进入角色扮演，或使用 [扮演xxx] 前缀时触发本 Skill。 基于 memory/*.md 中的历史对话（≥10条）提取说话风格，生成 persona/$user.json 档案。 支持：首次...

This skill appears coherent and implements exactly what it claims: scanning local memory/*.md messages, generating an LLM analysis prompt, and saving persona JSONs under persona/. Before installing/using it, consider: (1) memory/*.md contents will be included in prompts sent to your configured LLM — do not analyze logs containing passwords, private keys, medical/financial PII, or other sensitive text unless you trust the model provider and transmission. (2) Temporary files are written to /tmp (persona_analysis_prompt.txt and persona_meta.json); other local processes could read them. (3) Persona files and backups are written to persona/ and delete operations create .bak files — check those files for sensitive content before sharing. (4) The message-extraction heuristics may misclassify lines; inspect extracted samples (dry-run) before performing full analysis. If you accept those data-exposure tradeoffs, the skill is internally consistent.

Type: OpenClaw Skill Name: persona-creator Version: 1.0.0 The skill bundle is designed to analyze chat history in the 'memory/' directory to generate style profiles. It is classified as suspicious due to a high risk of command injection in SKILL.md, where unvalidated user-provided input (the 'name') is directly interpolated into shell command blocks (e.g., python3 ... --user "{用户输入的名字}"). While the Python scripts (analyze.py, save_persona.py) perform their stated functions and lack evidence of intentional malice or exfiltration, the architectural flaw in the instructions allows for potential Remote Code Execution (RCE) if a user provides a malicious name string.