xAI engineer uses Grok Bot with Cursor to merge 1,150 PRs in four days and build a pixel game

This latest development engineering retrospective publicly shared by Silicon Valley development goddess "Potato Sister" Lauren, I think is the most hardcore model of AI-assisted coding to date, bar none. Potato Goddess alone, in 4 days, merged about 1,150 PRs, and built a complete game that runs directly inside an X post, with a package size of only 124KB, running stably at 60 frames on a thousand-yuan phone with 500 monsters on screen at once, called tibo dog. Many people think using AI to write code is just throwing out a few prompts and then leaving it to luck, and the result is often increasingly messy code and more and more bugs to fix. But this top-tier goddess engineer showed the exact opposite stance: not letting AI run free, but using extremely strict engineering iron rules to keep AI firmly in line. Breaking down the 6 brutal engineering rules she wrote into the repository, each one redefines the upper limit of human-machine collaboration: 1️⃣ Turn every bug into a Lint static rule: every time you step on a pitfall in the code, the first reaction is not a one-off fix, but having the Bot write a Lint rule to completely seal off that way of writing, so it never happens again across the whole project. 2️⃣ Strictly forbid using code comments to explain logic: through Lint rules, writing comments is forcibly prohibited. If a piece of code needs explanation, the only allowed solution is to choose a better variable name, or add a unit test. 3️⃣ Use hot-path naming conventions to lock down memory garbage: all high-frequency code that runs every frame uses the .hot.ts suffix, and Lint rules clamp down hard: within that file, any object literals, closures, and new operations are strictly forbidden, so the main loop produces almost zero memory garbage, eliminating GC frame drops at the root. 4️⃣ Introduce TLA+ formal verification to find state machine vulnerabilities: perform strict formal mathematical proofs on cutscenes and core system state machines, draining all logical deadlocks before runtime. 5️⃣ Zero audio files and pure-code real-time chip synthesis: no external MP3 or audio packages are loaded at all; all music and sound effects are synthesized in real time by a self-developed Game Boy-style chip in the browser, and the background music can even seamlessly and smoothly vary between different terrains. 6️⃣ Fully deterministic input replay for anti-cheat and headless Bot balance tuning: game execution is completely deterministic, and the server determines whether a high score is cheating by replaying the player's input stream; at the same time, a headless AI Bot automatically runs levels overnight, automatically adjusting numerical difficulty based on survival rate. The harshest truth is: the more rigidly you write the engineering rules, the more explosively good the quality AI delivers. Maybe AI won't eliminate programmers, but top engineers who know how to drive AI with strict engineering standards are turning themselves into an entire independent studio with big-tech-level productivity.

View original

硅谷开发女神土豆姐 Lauren 最新公开的这篇开发工程复盘，我觉得是目前 AI 辅助编程最硬核的范本，没有之一， 土豆女神一个人、4 天时间、合并约 1150 个 PR，做出一款直接跑在 X 帖子内、包体仅 124KB、千元机稳跑 60 帧且带 500 个同屏怪物的完整游戏，叫tibo dog， 很多人以为用 AI 写代码就是随便给几句 Prompt 然后撞大运，结果往往是代码越来越乱、Bug 越修越多。 但这位顶级女神工程师展示了完全相反的姿态： 不是放任 AI 自由发挥，而是用极其严苛的工程铁律把 AI 治得服服帖帖。 拆解她写进仓库里的 6 条残暴工程规则， 每一条都在重新定义人机协作的上限： 1️⃣ 把每一个 Bug 变成 Lint 静态规则：每在代码里踩到一个坑，第一反应不是单点修复，而是让 Bot 写一条 Lint 规则把这种写法彻底封死，全项目永不再犯。 2️⃣ 严禁用代码注释来解释逻辑：通过 Lint 规则强制禁止写注释。如果一段代码需要解释，唯一允许的解法是换一个更好的变量名，或者补一个单元测试。 3️⃣ 用热路径命名规约锁死内存垃圾：每帧运行的高频代码全部以 .hot.ts 为后缀，Lint 规则死死卡住：该文件内严禁任何对象字面量、闭包和 new 操作，主循环几乎产生零内存垃圾，从根源上消灭 GC 掉帧。 4️⃣ 引入 TLA+ 形式化验证找状态机漏洞：对剧情过场与核心系统状态机做严格的形式化数学证明，在运行前把逻辑死锁全部排空。 5️⃣ 零音频文件与纯代码实时芯片合成：没有加载任何外部 MP3 或音频包，所有音乐和音效全由浏览器自研的 Game Boy 风格芯片实时合成，背景音乐甚至能在不同地形之间无缝平滑变奏。 6️⃣ 全确定性输入重放防作弊与无头 Bot 调平衡：游戏运行完全确定，服务器通过重放玩家的输入流判定高分是否作弊；同时用无头 AI Bot 自动通宵挂机跑关卡，根据存活率自动调节数值难度。 最狠的一条真相是： 你把工程规则写得越死，AI 交付出来的质量就越炸裂。 也许AI 不会淘汰程序员，但懂得用严格工程规范驱动 AI 的顶级工程师，正在把自己变成一整个拥有大厂产能的独立工作室。