Playbook：做 App

This is how I use Grok Bot to run a mobile game studio with a team of six bots (and growing). Each one has its own job, its own computer, and its own todo list, and together they build, ship, and improve my mobile game Rank'em, which just passed 1000 downloads across iOS and Android over the past week.

我用 Grok Bot 跑一个手游工作室，编制是六个 Bot（还在加）。每个有自己的岗位、自己的电脑、自己的待办，一起做、发、改我的手游 Rank'em。过去一周 iOS 和 Android 加起来刚过 1000 次下载。（六个 Bot 是作者自己的实验编制，不是产品硬限制。「自己的电脑」是作者的说法；正式文档写账号下所有 Bot 共用一台云电脑、各自屏幕。）

I started my career at Sensor Tower and Lion Studios (AppLovin) studying how mobile games get built, and the lesson was always the same: the game itself is maybe 15% of the job. Especially with Claude Code, Codex, or similar tools, anyone can one-shot a serviceable clone of multi-million-dollar games like Candy Crush or Wordle today. But then they're stuck with the other 85%.

我职业生涯从 Sensor Tower 和 Lion Studios（AppLovin）开始，研究手游怎么做出来。教训一直一样：游戏本身大概只占 15%。尤其有了 Claude Code、Codex 或同类工具，今天谁都能一锤子做出 Candy Crush、Wordle 那种千万级游戏的能玩仿品。然后就卡在剩下的 85%。

The six seats, as the author wrote them:

六个座位，按作者写的：

Grok Bot launched out of SpaceXAI this month. Each sidebar item is not a chat window, it's a long-lived agent. Think of each one as a DM with a coworker who has:

Grok Bot 这个月从 SpaceXAI 发出来。侧栏每一项不是聊天窗，是一个活得久的代理。把它想成跟同事的私信，这个同事有：

Job description (who it is): a system prompt that reads like a role — the lane it owns, and the work it refuses. Connections (what it can reach): the accounts it is signed into. Jira, Figma, GSuite, Salesforce, GitHub, or in my case eight ad and analytics tools. A computer in the cloud: its own machine, with a browser and a terminal, running whether or not yours is on. Routines (when it acts): standing work on a clock. Not a reminder you set, and not something you have to open the app to trigger. Skills (how it does a task): recorded playbooks. You do the job once while it watches, and it keeps the click-path. Handoffs (who it passes to): the ability to hand work to another bot directly, without routing through you. This is the one that makes it a team.

岗位说明（它是谁）：读起来像角色的系统提示——它管哪条车道，以及它拒绝的活。连接（能碰到什么）：它登进哪些账号。Jira、Figma、GSuite、Salesforce、GitHub，或作者这边八个广告和分析工具。云上的电脑：它自己的机器，有浏览器和终端，你的开不开它都跑。（「自己的机器」是作者的说法；正式文档写账号下所有 Bot 共用一台云电脑、各自屏幕。）Routines（何时动手）：按钟点的常设活。不是你设的提醒，也不是你打开 App 才触发。Skills（怎么干一件事）：录下来的操作手册。你干一遍它看着，点击路径它留下。Handoffs（传给谁）：直接把活交给另一个 Bot，不经过你。这一条才让它成为团队。

The last one is the one that matters. A bot that can hand work to another bot is a foundation for a team, and what really differentiates the Grok Bot experience from other knowledge-work products.

最后这条才是要紧的。一个能把活交给另一个 Bot 的 Bot，才是团队的地基，也是 Grok Bot 跟其他知识工作产品真正不一样的地方。

I wrote six seats out as job descriptions, the way a studio would, and handed one to each bot. At the beginning of each day each bot also had a particular task it was in charge of working on overnight, so it was ready for me in the morning.

我按工作室的写法，把六个座位写成岗位说明，一个座位交给一个 Bot。每天开头，每个 Bot 还背着一件隔夜要干的活，好让我早上起来就能用。

Mobile Orchestrator is the manager. It owns App Store and Google Play, and routes work between the other five. Analytics Agent owns paid acquisition and reads what players actually did. The only bot allowed to declare a finding. Creatives Agent turns findings into ad specs and renders the variants. It never buys the media itself. Rank'em Engineer owns the app and the backend. It takes a finding as a spec, not as a suggestion. GCS agent handles GKE, certs, deploys, and rollbacks. Nothing reaches players without going through it. Bug fix Agent sweeps Sentry overnight. It fixes what is obvious and escalates what is not.

Mobile Orchestrator 是经理。管 App Store 和 Google Play，在另外五个之间分活。Analytics Agent 管付费买量，读玩家实际做了什么。唯一有权宣布「结论」的 Bot。Creatives Agent 把结论变成广告规格并渲出变体。它自己从不买量。Rank'em Engineer 管 App 和后端。它把结论当规格接，不当建议。GCS agent 管 GKE、证书、部署、回滚。任何东西到玩家手里都得经过它。Bug fix Agent 隔夜扫 Sentry。明显的它修，不明显的它升级上报。

The feedback loop between analytics and gameplay is the most time-consuming part of running a live game, so it made the best candidate for turning into a bot first. Here is the rough breakdown of the configuration for the Analytics bot:

跑一款上线游戏，分析和玩法之间的反馈循环最耗时间，所以最适合先做成 Bot。下面是 Analytics Bot 配置的大致拆解。英文原文整段保留，后面逐条对照。

Another incredibly valuable feature of Grok Bot is the ability to "teach it" a skill by hitting the record button and doing something on Grok Bot's computer. I was unable to use the Meta Ads API because of verification issues, so rather than wait on Facebook support, I recorded myself uploading a creative through the Meta web UI once, and the bot learned and remembered the click-path for future creatives (it was uploading and testing at least one new creative per day).

Grok Bot 另一个非常有用的能力：按录制，在它的电脑上做一遍，就能「教」它一个 skill。Meta Ads API 我因为验证问题用不了，于是不等 Facebook 支持，自己在 Meta 网页上传播过一次素材，Bot 学会并记住了以后传素材的点击路径（当时每天至少上传并测试一条新素材）。

Here's an example of what now runs more autonomously as routines, but started as me asking my bots to do a little work as I went to bed. You may notice that I asked it to do something, and on this particular night the bot realized that my request should be a routine and went ahead and added that to its own remit so I never had to ask again.

有些现在更自主、当成 routine 在跑的事，一开始只是我睡觉前让 Bot 干一点小活。你会注意到：我让它做一件事，那天晚上 Bot 自己意识到这请求该变成 routine，就写进自己的职责范围，我不用再问第二次。（原文页上还有录像，这里只留这个意思。）

ROI is really the name of the game, and I plan to talk about this in more detail as I do the same work on revenue generation in the app. But two wins trace directly to the bots, resulting in a 15x improvement in cost per install and a 4x improvement in D7 retention.

ROI 才是这行的名字。我打算等在 App 里做同一套收入工作时再细讲。但有两场直接能追到 Bot 的胜：单次安装成本改善 15 倍，D7 留存改善 4 倍。

The bots analyzed the data, saw that people wanted more of a partner game than a party game, and kept revising the ad until installs cost 15x less. The same feedback loop produced both wins: the bots read the data, worked together as a team to form and test a hypothesis, and: (1) Swapped the party ad for the partner-cut ad after data showed people were playing as couples rather than large groups, taking installs from $15 to $1. (2) Shipped a bonus "hint" feature that shifted the whole score curve after proving that higher scores are correlated with better retention, improving D7 retention by about 4x.

Bot 分析数据，看出人更想要搭档游戏而不是派对游戏，于是一直改广告，直到安装成本降到原来的 1/15。同一条反馈循环产出两场胜：Bot 读数据，作为团队一起形成并验证假设。一，数据说明人是成对在玩、不是大团玩，于是派对广告换成搭档剪辑，安装从 15 美元降到 1 美元。二，证明更高分和更好留存相关之后，上线了额外「提示」功能，整条分数曲线都挪了，D7 留存大约改善 4 倍。

I've been pleasantly surprised by the progress so far, and just like building a team of humans, now that the MVP is ready it's time to scale up. As I continue to grow the team of bots, some of the empty seats are obvious, like setting up a revenue bot or adding more ad networks, and some will emerge as new problems and responsibilities arise.

目前的进展让我高兴得有点意外。跟人组团队一样，MVP 好了就该扩。继续加人时，有些空座位很明显，比如做一个收入 Bot、再加广告网络；有些会在新问题、新职责冒出来时才长出来。