Building an AI Development Workflow with Claude Code Plugins
A Juejin article describes an author-built system that organizes AI development workflows into layered Claude Code plugins for general processes, technology stacks, design tools, and API tools. The system uses on-demand skills, an initialization command that detects project details and generates configuration and document scaffolding, plus scripts, hooks, and CI checks to enforce required validations. Its example implementation uses Flutter, while the author says the underlying approach can apply to other technology stacks.
The article provides a concrete implementation pattern for teams seeking to make AI-assisted development processes portable, project-aware, and less dependent on the model remembering to perform mandatory checks.
导读
很多团队会把「AI 该怎么干活」写进项目的 CLAUDE.md 或一份很长的工作流文档：任务怎么分级、需求怎么拆、做完要跑哪些检查、怎么交付。用了一段时间后，我们碰到了三个问题：
- 复制走样：流程和项目特有的东西（技术栈、设计工具、接口平台、验证平台）混在一起，新项目只能整份复制再手工删改，改进也没法同步到其他项目。
- 只是文字：流程里说「需求要拆成任务卡」「中断要写交接记录」，但新项目里根本没有这些文件和目录，流程落不下来。
- 靠 AI 自觉：「改完代码要跑静态分析」写得再清楚，AI 也有漏掉的时候，而没人会发现。
我们的解法是把工作流做成一组 Claude Code 插件：流程写成按需触发的 skill，项目差异由一个 init 命令自动探测并生成配置和文档骨架，关键检查用脚本和钩子兜底。现在新项目接入只需两步：装插件、跑一次 /workflow-core:init。
#Claude Code#ai-dev-workflow#Flutter#Pixso#Apifox#Figma#React#Vue#Dart