The game is maybe 15% of a mobile studio. Anyone with a coding agent can one-shot a serviceable clone. The rest of the week is ads, attribution, store listings, crashes on one device, and a backend that survives a campaign.
Ryan Perry, writing on the official Grok Bot guides, used six named bots to run Rank'em, a game that had just crossed 1,000 downloads on iOS and Android in a week. This article is the Academy version: the seats, the first bot to build, the handoffs, and the job briefs we keep on Grok Bot engineering prompts and Grok Bot paid ads prompts.
Grok Bot here means the teammate with a cloud computer, not grok.com chat.
The 85% a coding agent will not sit in
Perry's table is the useful part. Steal the seats even if you never touch a game:
| Seat | The job you actually have |
|---|---|
| User acquisition | Lives in the ad manager. Decides what to spend tomorrow and which campaign dies today. |
| Performance creative | Ships new cuts every week because fatigue is measured in days. |
| Client engineering | The app, the progression, the content pipeline. |
| Backend and live ops | Saves, leaderboards, events, the reason progress survives a new phone. |
| Release management | Apple, Google, privacy labels, TestFlight, certs that expire at the worst time. |
| QA and incidents | The crash on one OS, the spike an hour after a release. |
One chat that "helps with the app" will mix all six. Six named bots will not, if you write the refuse-list into each description.
What Grok Bot is, in studio language
Each sidebar item is a long-lived agent:
- A job description: the lane, and the work it refuses
- Connections: the accounts it is signed into
- A computer that runs when your laptop is shut
- Routines on a clock
- Skills: a click-path you recorded once
- Handoffs: work passed to another bot without routing through you
Handoffs are the difference between a clever assistant and a team. A finding that cannot leave the analytics bot is a note. A finding that arrives on the engineer bot as a spec is work.
Six seats, rewritten without the game's proper nouns
Perry's roster, in jobs you can reuse:
- Orchestrator. Owns the stores. Routes work. Does not write app code.
- Analytics. The only bot allowed to declare a finding. Never writes creative. Never touches app code.
- Creatives. Turns a finding into an ad spec and variants. Never buys media.
- App engineer. Takes a finding as a spec, not a suggestion.
- Deploy. Certs, clusters, rollbacks. Nothing reaches players without it.
- Bugs. Sweeps the error tracker overnight. Fixes the obvious. Escalates the rest.
If you ship a SaaS instead of a game, keep the same six ideas: who may declare a finding, who may change production, who may spend.
Build analytics first
The loop between "what people did" and "what we shipped" is the slowest part of a live product. That made analytics the first bot worth configuring.
A job description in that shape looks like this. Fill the brackets. Do not let it spend.
You are the analytics seat for [APP]. You own paid acquisition reads and product reads. You are the only bot allowed to declare a finding. You never write creative and you never touch app code.
Sources: [ADS MANAGER], [ATTRIBUTION], [PRODUCT ANALYTICS], [STORES], [ERROR TRACKER].
If a tool is missing, ask me before working around it.
Every night: recap spend, installs, and cost per install. Name winners and losers.
Every morning: read [CORE FUNNEL]. Flag if the ad and the app describe different products.
Weekly: roll up 7-day and 30-day cost against store revenue.
Handoffs: winners and losers to Creatives. Product findings to the engineer as specs.
Check with me before anything that spends money.
Perry could not use the Meta ads API (verification). He recorded himself uploading a creative in the web UI once. The bot replayed the click-path. That is a Grok Bot skill: you do the job on its computer, with no microphone required if you want it that way, and you keep send and pay out of the recording. Our skills prompts are written for that.
What the official guide claimed, and what you should not copy blindly
Perry wrote that the same loop produced two wins: swapping a "party" ad for a "partner" cut after the data said people played as couples, taking installs from about $15 to about $1, and a hint feature after scores lined up with retention, about 4x on day-7.
Those numbers are his, on his game, in his window. Do not put them in a deck as if they are yours. Copy the loop: one bot may declare a finding, another may turn it into a spec, you still flip the campaign and you still ship the build.
What you still do
- Approve spend
- Submit the store listing
- Merge or reject the change that reaches players
- Take over the computer for 2FA
- Keep customer data off the shared machine unless you accept that every bot on the account can see it
Official docs are explicit: isolation is per user, not per bot. A second bot is not a security boundary. Put irreversible actions behind you. Grok Bot security prompts exist for that list.
Source
Field story: Grok Bot for mobile app development by Ryan Perry on x.ai. We kept the seats and the loop. We did not paste his studio config. Use our briefs, then change the tool names to yours.