Under the hood
The tools and architecture behind our 33 apps for iPhone, iPad and Apple Watch: AI to explore and build, Apple’s tools to test and ship, and a human in the loop at every key step.
Process flow
Five stages, with feedback loops in every one. Start with the simple view, or switch to the detailed one: tap any step or loop to see the tools, the people and agents involved, and what it produces. Ideas start as a single ball and build up to 10; tap one to follow it. Zoom in or out with the slider. The gold diamonds are the human-in-the-loop checkpoints.
Ideas from play, players and contributors, explored with ChatGPT and Grok.
Idea approvalAI agents help write the Swift and web code in Xcode and VS Code.
Code reviewSimulators, our own devices and TestFlight beta testers.
Go / no-goApp Store Connect, App Review and release.
Players play, leave reviews and send feedback.
Feedback loop: what players tell us sparks the next idea, and the cycle starts again. The gold tags are human-in-the-loop checkpoints.
Ideas flowing
01 Ideate
Ideas come from play, player feedback and our community contributors. We explore them in conversation with ChatGPT and Grok, and capture notes with Apple Intelligence.
Tools
Who
Paul and SophieOutput
A one-line idea worth exploring.
02 Ideate
Paul and Sophie decide which ideas go further. Nothing moves on until a human says yes.
Tools
A human decision
Who
Paul and SophieOutput
Keep, park, or kill kindly.
03 Ideate
The idea becomes a clear brief: what we’re making, who it’s for, the constraints and how we’ll know it works, refined across various GPT and Grok models.
Tools
Who
Paul and SophieOutput
A clear, reusable brief.
04 Feedback loop · Ideate
We try a prompt, look at what comes back, and refine the brief until the output is right, before any building starts.
Tools
Who
Paul and SophieOutput
A sharper brief.
05 Build
ChatGPT work mode and our Grok Bot agents work through the brief step by step: Onyx on features and fixes, Opal on icons and screens, Topaz on copy.
Tools
Who
Output
Features, screens and copy, ready for review.
06 Build
Apps are written in Swift; web experiences, including this site, in JavaScript, HTML and CSS; Python automates the tooling. Xcode compiles and signs each build, and every change goes to GitHub.
Tools
Who
PaulOutput
A signed build, with every change in GitHub.
07 Build
Apple frameworks such as CloudKit for iCloud data and StoreKit for in-app purchases, three.js for 3D on the web, and Google AdMob for ads in free apps.
Tools
Who
Output
Data, purchases, 3D and ads, wired in.
08 Build
A human reviews the code and plays the build. Approved builds fork into parallel testing; anything else goes back for rework.
Tools
Who
PaulOutput
Approved for testing, or sent back for rework.
09 Feedback loop · Build
If a review finds problems, the work goes back to the agents and into code for another pass, then it’s reviewed again.
Tools
Who
PaulOutput
Cleaner code, reviewed again.
10 Test
In parallel with alpha and beta testing, builds run in iOS device simulators across different iPhone and iPad screen sizes, catching layout and logic issues early.
Tools
Who
Output
Layout and logic issues caught early.
11 Test
At the same time as the simulator runs, alpha builds go onto our own test devices to check real-world feel, performance and hardware features.
Tools
Who
Output
Real-world feel and performance checked.
12 Test
Xcode uploads the build to App Store Connect and TestFlight delivers it to beta testers’ devices, including our community contributors.
Tools
Who
Output
Tester feedback for the final polish.
13 Feedback loop · Test
Layout or logic issues found in the simulators go straight back to Build to be fixed, then the build runs again.
Tools
Who
Output
A fixed build, run again.
14 Feedback loop · Test
Issues found on alpha devices or reported by TestFlight testers go back to Build, and a new build goes out for another round.
Tools
Who
Output
A new build for testers.
15 Ship
The parallel test branches meet here. When simulator, alpha and beta testing all check out, a human decides: only a human can press go on a release.
Tools
A human decision
Who
Paul and SophieOutput
Go, or back for another pass.
16 Ship
The store listing, screenshots and creative are set up in App Store Connect and the build is submitted for App Review. Once approved, it’s released on the App Store.
Tools
Who
Output
Released on the App Store.
17 Feedback loop · Ship
If App Review asks for changes, we fix what’s needed, upload a new build and resubmit, through the go/no-go gate again.
Tools
Who
PaulOutput
Approved for release.
18 Ship
New features can ship switched off inside a live app. We turn them on for a small group or our testers first, measure how they go, then switch them on progressively, or roll them back and send them back to testing, without a new release.
Tools
Who
Paul and SophieOutput
A feature switched on for everyone, or rolled back.
19 Enjoy
Players find and download our apps on the App Store: 33 apps for iPhone, iPad and Apple Watch, with 21k+ downloads so far.
Tools
Who
Output
Downloads.
20 Enjoy
People use and play our apps and games. Free apps may show ads from Google AdMob, any purchases go through Apple with StoreKit, and features piloted with feature toggles reach players here.
Tools
Who
Output
Fun, usefulness and feedback.
21 Enjoy
Feedback arrives as App Store reviews, support requests through our Support page, and usage insights from App Store Connect and AdMob.
Tools
Who
Paul and SophieOutput
What to fix, improve or make next.
22 Feedback loop · Enjoy
Everything we learn from players flows back to Ideate, shaping fixes, new features and the next app. Feedback also tells feature toggles what to switch on next.
Tools
Who
Paul and SophieOutput
The next idea.
Architecture
Every app moves left to right through the same stack, then what we learn loops back to the start.
Paul and Sophie steer set direction, review the work and approve every hand-off
HITL01 Ideate
02 Build
Code & frameworks
03 Test
04 Ship
Live in the apps
05 Enjoy
Various GPT and Grok models power the AI tools in every layer
Feedback loop
Feedback loops run inside every layer. The big one carries player reviews, support requests, App Store Connect and AdMob insights and TestFlight feedback back into the next idea.
Architecture overview. Paul and Sophie steer the whole pipeline and approve every hand-off. Layer 1, Ideate: prompt engineering, ChatGPT chat mode, Grok and Apple Intelligence. Layer 2, Build: ChatGPT work mode, Grok Bot with the AI agents Onyx, Topaz and Opal, various GPT and Grok models, VS Code, Xcode and GitHub, with code in Swift, JavaScript, HTML, CSS and Python, and the frameworks and services three.js, CloudKit, StoreKit and Google AdMob. Layer 3, Test: iOS device simulators running in parallel with alpha and beta testing on test devices and TestFlight, with Jasper and community contributors Kristen, Alice and Mary. Layer 4, Ship: App Store Connect and feature toggles, with StoreKit purchases and Google AdMob ads live in the apps. Layer 5, Enjoy: the App Store, where players download, use and play our apps and give feedback through reviews, support requests and usage and AdMob insights. Various GPT and Grok models power the AI tools across every layer. Feedback loops run within every layer, and a big loop carries player feedback, App Store Connect and AdMob insights and TestFlight feedback back to the next idea.
Languages & frameworks
Five languages, three frameworks and libraries, and one ads service, each matched to where it runs.
Language
Apple frameworks
Services
In-app purchases go through StoreKit and are processed by Apple. Free apps may show ads from Google AdMob.
Languages
Library
Hosting
Interactive pages and 3D in the browser, plus this website, served by GitHub Pages.
Language
Runs alongside
Scripts that build this site and automate App Store Connect work.
The pipeline
What happens at each step, and the tools that do the work.
Ideas come from play, player feedback and our community contributors. We explore them in conversation with ChatGPT and Grok, and capture and tidy notes with Apple Intelligence.
HITLIdea approval Paul and Sophie decide which ideas go further.
The idea becomes a clear brief: what we’re making, who it’s for, the constraints and how we’ll know it works. Good prompts are reused and refined across various GPT and Grok models.
ChatGPT work mode and our Grok Bot agents take the brief and work through it step by step: Onyx on features and fixes, Opal on icons and screens, Topaz on copy.
Apps are written in Swift; web experiences, including this site, in JavaScript, HTML and CSS; and Python scripts automate the tooling. Code is edited in VS Code and Xcode. Xcode compiles and signs each build, and every change is committed to GitHub so it can be reviewed and rolled back.
HITLCode & build review A human reviews the code and plays the build before testing.
Apple frameworks such as CloudKit for iCloud data and StoreKit for in-app purchases, three.js for 3D on the web, and Google AdMob for ads in free apps.
In parallel with alpha and beta testing, builds run in iOS device simulators across different iPhone and iPad screen sizes, catching layout and logic issues early. Issues go straight back to Build.
At the same time as the simulator runs, alpha builds go onto our own test devices to check real-world feel, performance and hardware features.
Xcode uploads the build to App Store Connect and TestFlight delivers it to beta testers’ devices, including our community contributors Kristen, Alice and Mary. Their feedback shapes the final polish, and any issues loop back to Build.
The store listing, screenshots and creative are set up in App Store Connect, the build is submitted for App Review (if it asks for changes, we fix and resubmit), and once approved it’s released on the App Store.
HITLRelease go/no-go When simulator, alpha and beta testing all check out, only a human can press go on a release.
New features can ship switched off inside a live app. We turn them on for a small group or our testers first, measure how they go, then switch them on progressively, or roll them back without a new release.
Players download, use and play our apps, and tell us what they think through App Store reviews and our Support page. App Store Connect analytics, AdMob reports and tester feedback show what’s working, and it all feeds straight into the next idea.
The toolchain
24 tools, languages and frameworks, each with one clear job.
01 Ideate
Turning a rough idea into a clear, reusable brief, with goals, audience, constraints and checks, so AI tools give focused, consistent results.
Conversational brainstorming, research and problem-solving, from first concepts to tricky bugs.
A second conversational partner for ideas, research and fresh perspectives on a problem.
Apple’s built-in tools on our own devices for capturing, summarising and polishing notes and text.
02 Build
Agent mode that takes a brief and works through multi-step tasks, then reports back for review.
Home of our AI agents, Onyx, Topaz, Opal and Jasper, who carry out engineering, marketing, design and QA tasks.
Different models for different jobs, chosen per task, from quick drafts to deeper reasoning.
A lightweight editor for code, scripts, web pages and config.
Apple’s developer tool: builds, signs and archives our apps, runs the simulators and uploads builds.
Source control for every change: versioned, reviewed and backed up. It also hosts this website.
03 Test
Run each build in Xcode on virtual iPhones and iPads of different sizes before it touches a real device.
Real hardware for testing: alpha builds on our own devices first, then beta builds on testers’ devices.
Apple’s beta service: installs pre-release builds on testers’ devices and collects their feedback.
04 Ship
Where each app is managed: builds, store listing, screenshots, pricing, App Review, release and analytics.
Switches that let a new feature ship turned off in a live app, be piloted with a small group, then be turned on gradually or rolled back without a new release.
+ Languages, frameworks & services
Apple’s programming language, used for our apps on iPhone, iPad and Apple Watch.
Brings our web experiences, and this website, to life with interactivity.
The structure and content of our web pages, including this site.
The look, layout and motion of our web pages, from colour to responsive design.
Tooling and automation, such as this site’s generator and our App Store Connect scripts.
A JavaScript library for real-time 3D graphics in the browser.
Apple’s framework for storing app data in iCloud and syncing it across a player’s devices.
Apple’s framework for in-app purchases and subscriptions. Apple processes every payment.
Google’s mobile ads platform, used to show ads in our free apps, with reports on how they perform.
Browse all 33 apps, or see how the team works together.