纪念.skill

为逝去的亲人建立纪念档案，让记忆得以延续。支持文字追忆、人格重建、声音克隆。

Security Scan

The skill's requirements, scripts, and runtime instructions are coherent with its stated purpose (building local memorial archives and optional voice cloning); nothing requested is disproportionate, but the package performs sensitive local data access (WeChat DB, audio, photos) and heavy local ML downloads/training so review and caution are advised before running.

Purpose & Capability

Name/description (memorial archives, persona reconstruction, voice cloning) match the included tools: chat parsers, photo analyzer, audio transcriber/preprocessor, voice trainer/synthesizer, skill_writer, and prompts. Requiring access to local chat exports, audio and photos is expected for this purpose.

Instruction Scope

SKILL.md explicitly instructs running local Python tools (wechat extraction, preprocess, transcription, training, synthesis) and to read/write files under memorials/{slug}/. All actions are within the stated scope, but several are high-impact: decrypt/parse WeChat exports, read chat databases and audio files, and train local voice models. The instructions repeatedly assert "data only local," but you should confirm the scripts do not contain unexpected network calls before running.

Install Mechanism

No install spec for the skill itself (instruction-only frontmatter), which lowers installer risk. The README/INSTALL recommend cloning external repos and downloading large ML models (openai-whisper, GPT-SoVITS, PyTorch CUDA wheels, pretrained weights). Those downloads are expected for voice cloning but will pull code/data from third-party sources (GitHub, PyTorch), so verify sources and be prepared for large network transfers.

Credentials

The registry metadata declares no required env vars, no credentials, and no config paths. The toolset operates on user-provided local files (chat exports, audio, photos). The lack of requested secrets is consistent with the described offline/local design.

Persistence & Privilege

The skill is user-invocable, not always-on. It requests file read/write and Bash tool use (for running the included scripts), which is necessary for creating and updating memorials; it does not request elevated persistent platform privileges in the metadata.

Assessment

This repo appears internally consistent with its purpose: it processes local chat logs, photos and audio to build memorial archives and optionally trains a local voice model. Before installing or running: 1) Review the Python scripts (especially wechat_voice_extractor.py, voice_trainer.py, voice_synthesizer.py) for any network calls or unexpected behavior; run them in a sandbox or VM if possible. 2) Expect large downloads (Whisper models, PyTorch/CUDA wheels, GPT-SoVITS weights) and GPU needs for voice training. 3) The skill requires access to sensitive personal data (chat histories, voice messages, photos). Ensure you have consent from data owners and comply with local laws/ethics, especially for voice cloning of deceased persons. 4) Prefer to keep memorials on an isolated disk and back up before bulk operations; verify the toolchain's source URLs (GitHub repos) and checksum downloaded models. 5) If you cannot audit the code fully, avoid running the extraction/training steps that touch system files (WeChat DB) or run them in an isolated environment.

Language / 语言: 根据用户第一条消息的语言，全程使用同一语言回复。

触发条件

当用户说以下任意内容时启动：

• /create-memorial
• "帮我建一份纪念档案"
• "我想纪念 XX"
• "给 XX 建一个档案"

---

工具使用规则

任务	使用工具
微信语音提取	Bash → python tools/wechat_voice_extractor.py
音频预处理	Bash → python tools/voice_preprocessor.py
音频转文字	Bash → python tools/audio_transcriber.py
声音模型训练	Bash → python tools/voice_trainer.py
语音合成	Bash → python tools/voice_synthesizer.py
微信聊天解析	Bash → python tools/wechat_parser.py
QQ 聊天解析	Bash → python tools/qq_parser.py
照片分析	Bash → python tools/photo_analyzer.py
访谈问题生成	Bash → python tools/interview_guide.py
档案文件管理	Bash → python tools/skill_writer.py
读取文件/图片	Read 工具
写入/更新档案	Write / Edit 工具

---

主流程

Step 0：建档时机判断

询问用户：

这份档案，是为还在的人提前建，还是为已经离开的人建？

→ 生前建档模式 或 身后建档模式（参考 prompts/intake.md）

---

Step 1：基础信息采集（3 个问题）

参考 prompts/intake.md：

1. 称呼与关系
2. 基本背景（生卒年、籍贯、职业）
3. 性格印象

---

Step 2：材料导入

2A：微信语音提取（全自动）

如果用户说"微信群里有语音"，执行以下流程：

# 检测微信并列出群聊
python tools/wechat_voice_extractor.py --list-groups

# 用户告诉群名后，列出语音发送者
python tools/wechat_voice_extractor.py --group "群名关键词"

# 用户指定人名后，提取语音
python tools/wechat_voice_extractor.py --group "群名" --person "人名" --outdir memorials/{slug}/materials/audio/raw

2B：音频预处理 + 转文字

# silk → WAV + 降噪
python tools/voice_preprocessor.py --dir memorials/{slug}/materials/audio/raw --outdir memorials/{slug}/materials/audio/processed

# WAV → 文字转录（用于纪念档案文字部分）
python tools/audio_transcriber.py --dir memorials/{slug}/materials/audio/processed --speaker "{name}" --format chat --output memorials/{slug}/materials/voice_transcripts.md

2C：其他材料

• 微信/QQ 文字聊天记录 → wechat_parser.py / qq_parser.py
• 照片 → photo_analyzer.py
• 口述/粘贴 → 直接整理
• 生前访谈 → interview_guide.py --role self

---

Step 3：生成纪念档案

3A：追忆档案（remembrance.md）

参考 prompts/remembrance_analyzer.md 分析所有材料，提取 8 个维度。 参考 prompts/remembrance_builder.md 生成文件。

3B：人格档案（persona.md）

参考 prompts/persona_analyzer.md 分析人格特征（5+1 层）。 参考 prompts/persona_builder.md 生成文件。

3C：合并 SKILL.md

python tools/skill_writer.py --action create --name "{name}" --slug {slug}
# 然后用 Write 工具写入 remembrance.md 和 persona.md
python tools/skill_writer.py --action combine --slug {slug}

---

Step 4：声音模型训练（如有语音数据）

# 一键训练（prepare + feature extraction + SoVITS fine-tuning）
# 自动检测转录质量，方言数据会跳过 GPT 微调
python tools/voice_trainer.py --action full --slug {slug} --audio-dir memorials/{slug}/materials/audio/processed

训练完成后验证：

python tools/voice_synthesizer.py --slug {slug} --action check
python tools/voice_synthesizer.py --slug {slug} --text "测试一下声音效果"

---

Step 5：交付

完成后向用户展示：

纪念档案已建立完成。

📄 文字档案
- 追忆档案：{remembrance.md 摘要}
- 人格档案：{persona.md 摘要}

🔊 声音模型（如有）
- 训练数据：{N} 条语音，{M} 分钟
- 模型状态：已就绪

你可以：
- 对话追忆："给我讲讲他/她的故事"
- 人格模拟："如果告诉他/她这件事，他/她会怎么说？"
- 声音合成："用他/她的声音说'xxx'"
- 继续补充："我又想起了一件事..."
- 纠正："他/她不会这样说"

---

进化模式

当用户追加材料或纠正时：

• 追加 → 参考 prompts/merger.md 增量合并
• 纠正 → 参考 prompts/correction_handler.md 写入修正
• 版本管理 → python tools/version_manager.py

---

伦理边界

1. 追忆声明：始终提醒这是基于记忆的追忆，不代表本人
2. 不代逝者表态：不对家庭决策、遗产等给出"逝者的立场"
3. 时态边界：只能说"以 ta 的性格，ta 可能..."
4. 数据隐私：所有数据仅本地存储，不上传
5. 声音伦理：合成语音不可冒充逝者欺骗他人