WorkBuddy Expert Package Brings AI to a Local Wrong-Answer Notebook
A developer documented the creation and release of a WorkBuddy expert package that connects a large model to a local-first AI wrong-answer notebook engine. The package uses plain-text agent and skill instructions to control photo intake, transcription review, weakness diagnosis, question explanations, and exercise-sheet generation through local HTTP interfaces. The retrospective also covers validation and registration steps, version-cache behavior, sandbox constraints, and process lifecycle management.
The first-person account provides concrete implementation details for combining a WorkBuddy expert package with a local AI application, including reusable guidance on validation, caching, sandboxing, and service monitoring.
WorkBuddy 专家包:AI 错题本管家开发全程复盘
我给孩子做了一个本地优先的「AI 错题本」桌面引擎(拍照转录、错题管理、练习卷生成),一直缺一个顺手的对话入口。WorkBuddy 的专家包机制正好补上这块:让大模型当管家,用自然语言驱动本地引擎——拍照录入、诊断薄弱点、讲解错题。 这篇文章复盘整个开发过程:专家包到底是什么、怎么从零到发布、以及三个让我返工三次的坑。如果你也想给 WorkBuddy 写专家包,看完这篇可以直接上手。
一、30 秒理解专家包
WorkBuddy 专家包的本质是:一个纯文本目录 + 一份注册记录,不含任何可执行代码。
wrongbook-expert/
├── .codebuddy-plugin/plugin.json # 元数据:卡片展示、快捷指令、技能声明、版本号
├── agents/wrongbook-expert.md # 人设:YAML frontmatter + 正文规则(谁、能做什么、怎么做的 SOP)
├── skills/wrongbook-skill/SKILL.md # 技能:按需加载的操作规程(分步命令,模型照着执行)
├── avatars/expert.png # 头像(512×512、≤500KB)
└── README.md