Chinese Bedtime Story Generator

生成多角色中文睡前故事并合成语音。适用于用户想为孩子定制个性化睡前故事的场景：根据孩子姓名、年龄和兴趣，由 LLM 创建完整世界观和角色，生成分段故事文本（每段标注说话人），再由 TTS 以不同音色合成旁白、主角、小伙伴和长者的语音，最终拼接为完整 MP3 音频文件。支持连载模式（`--continue`）在多次...

Security Scan

The skill's code and instructions broadly match the described LLM+TTS bedtime-story purpose, but there are important incoherences and undisclosed requirements (missing declared env vars, unknown default endpoints, and a questionable MP3 concatenation approach) that warrant caution before installing.

Purpose & Capability

The name/description claim LLM + TTS multi-role story generation — the code indeed calls an LLM and a TTS service and saves story_state and MP3 outputs, so purpose and capability are generally aligned. However, the registry metadata lists no required environment variables or primary credential while the code requires STORY_LLM_API_KEY (or IME_MODEL_API_KEY) and STORY_TTS_API_KEY (or SENSEAUDIO_API_KEY). That mismatch between declared requirements and actual code is an incoherence.

Instruction Scope

SKILL.md and scripts limit behavior to generating story text, calling an LLM, calling a TTS API, persisting story_state.json and output files. They do not request unrelated files or broad system reads. Two issues to note: (1) the instructions/code claim MP3 frames can be 'directly binary appended' to produce a single MP3 without ffmpeg — this is generally unreliable and likely to produce invalid/uncertain audio files, not a security risk but an implementation bug; (2) prompts/behavior rely on saving and loading story_state.json for continuation, which is expected but means persisted state contains child name/age/interests and story summaries.

Install Mechanism

No install spec; this is effectively instruction + Python script. A requirements.txt lists openai, requests, python-dotenv which is proportionate to contacting LLM/TTS and loading .env. No arbitrary remote downloads or install-time code execution were specified.

Credentials

Although TTS/LLM API keys are appropriate for the stated functionality, the skill metadata declares no required env vars while the code requires STORY_LLM_API_KEY/IME_MODEL_API_KEY and STORY_TTS_API_KEY/SENSEAUDIO_API_KEY at runtime. Defaults point to third-party endpoints (models.audiozen.cn and api.senseaudio.cn). This is problematic because (a) the registry did not disclose credential needs, and (b) the default endpoints are non-obvious third-party services — you should verify those providers and their privacy/policy before supplying keys. The code will raise and exit if keys are missing when those functions run.

Persistence & Privilege

The skill does not request elevated privileges or permanent installation. It writes output and a story_state.json into an 'outputs' directory under the skill; always:false and no modifications to other skills or system-wide settings. This file-based persistence is expected for a continuation feature.

What to consider before installing

This skill mostly does what it says (use an LLM + TTS to build MP3 bedtime stories), but there are some red flags to check before installing or providing secrets: 1) The registry/metadata claims no environment variables, but the script requires STORY_LLM_API_KEY (or IME_MODEL_API_KEY) and STORY_TTS_API_KEY (or SENSEAUDIO_API_KEY). Do not provide real API keys until you verify the endpoints and provider reputation. 2) The code defaults to models.audiozen.cn and api.senseaudio.cn — confirm these are legitimate services, read their privacy/policy, and consider using your own known LLM/TTS endpoints by setting the STORY_LLM_BASE_URL / STORY_TTS_URL env vars. 3) Test the skill in a sandbox or with dummy keys and use --no-tts to exercise only text-generation first. 4) Be aware the MP3 concatenation is implemented by naive binary append (the README asserts 'no ffmpeg needed'); that may produce broken files — test output integrity before trusting it for playback. 5) Inspect the code yourself (or ask the author) if you plan to store real child data; story_state.json will persist child_name/age/interests and summaries locally. If you need help verifying the remote providers or altering the script to use a different TTS/LLM endpoint, consider that before entering secrets.

中文睡前故事生成器

适用范围

此 Skill 用于生成可播放的多角色中文睡前故事音频。

能力边界：

• 依赖 LLM 生成故事文本，TTS 合成多角色语音
• 支持自定义孩子姓名、年龄、兴趣爱好
• 支持连载模式，跨会话保持世界观和情节连贯
• 输出完整 MP3 音频 + 故事文本 + 状态文件

不做：

• 实时语音对话或 ASR 识别
• 视频或动画生成
• 英文故事

默认配置

• child_name: 小朋友
• age: 5
• interests: 冒险,动物
• episodes: 1（单集）

音色分配

角色	voice_id	speed	pitch
旁白 narrator	male_0004_a	0.9	0
主角 protagonist	child_0001_a	1.0	0
小伙伴 sidekick	child_0001_b	1.0	0
长者 elder	male_0018_a	0.85	-2

工作流

1. 初始化

   • 读取 --child-name、--age、--interests 参数
   • 若 --continue，从 story_state.json 加载已有世界观和角色

2. 世界观与角色创建（首次运行）

   • LLM 生成世界名称、背景设定、4个角色（旁白/主角/小伙伴/长者）
   • 主角名称默认使用孩子姓名

3. 故事生成

   • LLM 生成 12-20 段 segments，每段标注 speaker 和 text
   • 连载模式下传入上集摘要，保持情节连贯

4. TTS 多角色合成

   • 逐段根据 speaker 选择对应音色参数
   • 合成每段音频

5. 音频拼接

   • MP3 帧独立可解码，直接二进制追加拼接
   • 无需 ffmpeg

6. 保存输出

   • story_state.json：世界观+角色+情节摘要
   • story_ep{N}.txt：故事文本
   • story_ep{N}.mp3：完整音频

Prompt 模块

详见 references/prompts_cn.md。

数据结构

详见 references/state_schema_cn.md。

直接运行

pip install -r requirements.txt

# 首次生成
python scripts/run_story.py --child-name "小明" --age 5 --interests "恐龙,太空"

# 连载续写
python scripts/run_story.py --continue

# 不调用 TTS，仅输出文本
python scripts/run_story.py --child-name "小红" --age 7 --interests "魔法,精灵" --no-tts

# 生成多集
python scripts/run_story.py --child-name "小明" --age 5 --interests "恐龙,太空" --episodes 3

环境变量参考：.env.example

接口约定：

• LLM 读取 STORY_LLM_API_KEY，回退到 IME_MODEL_API_KEY
• TTS 读取 STORY_TTS_API_KEY，回退到 SENSEAUDIO_API_KEY