Build your own with
vibe coding.用 vibe coding 构建
属于自己的.
A college student hooked on Vibe Coding, turning ideas into working software. From full-stack web apps to macOS native tools to Android apps — every project is a new experiment.一个热爱 Vibe Coding 的大学生，用代码把想法变成现实。从全栈 Web 应用到 macOS 原生工具，再到 Android 应用，每个项目都是一次新的实验。
A student who believes code is expression.一个相信 代码即表达 的大学生.
I'm CityGenius, a college student who loves Vibe Coding — turning sparks of ideas into working products fast. I learn best by building; every project is a zero-to-one exploration.我是 CityGenius，一名在校大学生，热爱 Vibe Coding，用代码将灵感快速变为可用的产品。我相信最好的学习方式是动手做，每个项目都是一次从 0 到 1 的探索。
From Python full-stack to Swift native apps, from Kotlin Android to AI model integration — I enjoy jumping between stacks. This blog documents my builds, my thinking, and the pits I fell into along the way.从 Python 全栈开发到 Swift 原生应用，从 Kotlin Android 开发到 AI 模型集成，我享受在不同技术栈之间切换的乐趣。这个博客记录我的开发历程、技术思考和踩坑经验。
from Android to AI —
wiring different stacks
into one evolving
personal lab.从 Python 到 Swift，
从 Android 到 AI，
把不同技术栈
连接成一个持续演化的
个人实验室。
Every project is a zero-to-one exploration.每个项目，都是一次 从零到一 的探索.
From research tools to productivity apps — solving real problems with code.从学术研究工具到生产力应用，用代码解决真实问题。
Scholar AI
A research companion that helps scholars search, organize, and analyze papers with AI. Multilingual translation and smart summaries included.学术研究辅助工具，利用 AI 帮助研究者高效检索、整理和分析学术文献。支持多语言翻译和智能摘要。
macOS Native ToolsmacOS 原生工具
Efficiency tools built for macOS — native APIs, buttery interactions, deep system integration.为 macOS 打造的效率工具，利用原生 API 提供流畅体验，深度整合系统功能，追求交互细节。
Android App CollectionAndroid 应用集
A collection of Android apps: file conversion, on-device AI imaging, multilingual translation — pretty and practical.多个 Android 应用项目，覆盖文件格式转换、AI 图片处理、多语言翻译等功能，兼顾美感与实用性。
Stack技术栈
Python / FastAPI / React / Swift / AppKit / Kotlin / Jetpack Compose / TFLite / RAG.
Distilled from practice, discovered in code.
Ship a Full-Stack AI App in 3 Days with <em>Vibe Coding</em>用 <em>Vibe Coding</em> 三天搞定一个全栈 AI 应用
From idea to launch: building scholar-ai. How AI-assisted coding helped ship a FastAPI backend and a React frontend fast, plus real-world lessons from implementing RAG.
从想法到上线，记录 scholar-ai 的开发过程。如何利用 AI 辅助编码，快速搭建 FastAPI 后端和 React 前端，以及 RAG 检索增强生成的实际落地经验。
File Format Conversion on AndroidAndroid 文件格式转换方案
How to do quality file format conversion on Android. The on-device vs. cloud API tradeoff, plus UI practice with Jetpack Compose.
如何在 Android 上实现高质量的文件格式转换。本地处理 vs 云端 API 的权衡，以及 Jetpack Compose 的 UI 实践。
Write Agent Skills Like <em>Reusable Lesson Plans</em>把 Agent Skill 当成<em>可复用教案</em>来写
The biggest time sink in student AI projects isn't a weak model — it's re-explaining the same workflow every time. Writing workflows as Agent Skills is like writing experiment steps into a reusable lesson plan.
学生做 AI 项目时最容易浪费时间的，不是模型不够强，而是每次都要重新解释同一套流程。把工作流写成 Agent Skill，就像把实验步骤写进可复用的教案。
On-Device AI Models in Practice本地 AI 模型实战
Exploring what local AI models can do on mobile. TFLite image processing, offline translation model integration, and performance optimization.
探索在移动端运行本地 AI 模型的可能性。TFLite 图片处理、离线翻译模型的集成，以及性能优化技巧。
d-token: Before Context Leaves Your Machined-token：上下文离开本机之前
Compress redundancy, make routing explicit, and leave a receipt — before the request goes out. One real request sent 33,705 fewer tokens.
在请求发出前压缩冗余、写清路由、留下回执。一次真实请求少送 33,705 token。
agent-mcp: Turning Agent CLIs Into One Work Poolagent-mcp：把 Agent CLI 收成一个工作池
Opening more terminals is easy; staying in control is hard. The main agent only splits and merges — dispatch, timeouts, and resume go to the control plane, and each CLI gets matched to the task.
多开终端不难，难的是可控。主 Agent 只拆解汇合；派发、超时、续接交给控制面，CLI 按任务匹配。
From Desktop App to Claude Skill: Distilling a Design System's Reusable Logic从桌面应用到 Claude Skill：如何提炼设计系统的可复用逻辑
Using Open Design as an example: how to break down an app with 149 design systems, 110 templates, and 11 craft specs into a clean Claude Code Skill. Core ideas: three-axis composition, six-part assembly, advisor-style interaction.
以 Open Design 为例，讲解如何将一个包含 149 个设计系统、110 个模板、11 套工艺规范的应用，拆解为一个结构清晰的 Claude Code Skill。核心思路：三轴组合、六段拼装、顾问式交互。
EduEvidence: No Evidence, No New Study Designeduevidence：没有证据，就不设计新研究
Retrieved snippets are not evidence. A nine-stage protocol + evidence graph that outputs ADOPT, PILOT, REJECT, or INSUFFICIENT EVIDENCE.
检索片段不是证据。九阶段协议 + 证据图，给出 ADOPT、试点、驳回或证据不足。
ompweb: A Local Workbench for ompompweb：给 omp 一张本地工作台
No rewriting the agent — just read the local session files. Session tree, terminal, MCP, and worktrees, all on one desk.
不重写 Agent，只读本地会话文件。会话树、终端、MCP、worktree 都摊在一张桌子上。
SwiftUI vs. AppKit: Knowing When to SwitchSwiftUI 与 AppKit 的取舍
Lessons and pitfalls from building macOS tools. The conveniences and limits of SwiftUI, when to go back to AppKit, and how to do on-device inference with Core ML.
在开发 macOS 工具时遇到的坑和经验。SwiftUI 的便利与局限，什么时候该回归 AppKit，以及如何利用 Core ML 做本地推理。
From projects to method.从 项目 到 方法.
One feature piece, plus further reading.精选一篇主文，和两条延伸阅读。
Feature主文
Shipping a full-stack AI app in 3 days with Vibe Coding — the full log from spec to MVP.用 Vibe Coding 三天搞定一个全栈 AI 应用，从需求描述到 MVP 上线的完整记录。
Takeaway心得
Don't describe the whole page at once. Break it into small components with tight constraints — output quality jumps.不要一次描述整个页面。拆成小组件、逐个约束，AI 的输出质量会明显提升。
Limits限制
Complex business logic, perf-critical paths, deep mechanism understanding — still needs a human in the loop.复杂业务逻辑、性能关键路径、底层机制理解，仍然需要开发者亲自把关与修正。
Next延伸
Carrying this approach into macOS native tools, Android apps, and on-device AI.继续把这种开发方法迁移到 macOS 原生工具、Android 应用和本地 AI 集成场景里。
@37chengshan
Into my projects? Want to trade Vibe Coding notes? Or just talk tech? This is the front door.对我的项目感兴趣？想交流 Vibe Coding 的经验？或者只是想聊聊技术？这里是最直接的入口。
hello@citygenius.dev
For collabs, suggestions, or a deeper thread on any project — email gives us more room.如果你想合作、提建议，或者只是想继续深入聊某个项目，邮件会更适合展开。
Turn ideas into work.把 想法 继续做成 作品.
Find me on GitHub or by email — projects, tech, anything.欢迎从 GitHub 或邮件找我，聊项目、聊技术都行。
