Maker Lab in twelve days: an AIApplications workflow for building an iPhone game with AI agents

AI is often blamed at the moment for job loss, schoolwork, or the decline of online content. A more useful AIApplications angle is a practical workflow: Maker Lab, a free iPhone game on the App Store, moved from an empty folder to a finished build in twelve days during a recent staycation without hand-coded scripts. The lesson is not “AI replaces makers”; it is that a describe, play, complain, repeat loop can be reused for small products.

The Maker Lab main menu

Why the application matters more than the hype

The point is the use case: an AI coding agent helped turn an idea into a playable mobile app. That makes Maker Lab a clean example for office workers, product managers, operations leads, and digital project teams who do not need model internals. They only need a repeatable way to ask for an output, test it, and correct it.

The scenario: a makerspace game that tests process management

If players know Overcooked, the shape is easy to explain. Here, the player runs a makerspace. At the hatch, customers ask for a keyring, an engraved coaster, or a drone frame. The workspace uses 3D printers, a laser cutter, a CNC mill, a couple of hand benches, and exactly one pair of hands. Stock must be fetched, machines loaded, parts collected, and orders delivered before a customer leaves.

Level one: a scruffy garage, a bare bulb, and one 3D printer

The first screen is deliberately small: Level one begins with one 3D printer and a tidy constraint set. The progression is visual and motivational. Ten levels become ten rooms: junk disappears, shelving arrives, and a rusty bike is restored as a lit centrepiece. Later failures are part of the workflow training: from level five, machines can jam; an ignored jam can catch fire. From level nine, power can trip, and the breaker is far away. The hand benches still work because hands do not need electricity.

Fetch the extinguisher to put out the fire

The difficulty system also protects replay value. A bot must verify each challenge can be cleared, while keeping three stars out of reach on the last three levels, so a human still has something to chase.

The AI build workflow, step by step

The developer used Claude Fable, an Anthropic top-tier model, through Claude Code as a coding agent. Because the agent had access to the project folder, it could build the app, run tests, open it in iOS Simulator, and inspect screenshots. The human acted like a producer: describe, play, complain, adjust.

1. Describe the experience, not the code. A few paragraphs gave the first prompt, and the first commit produced a playable Level one.
2. Play it on the target device after every change. Most of the 144 commits came from phone testing, using plain corrections such as, “I can’t tell which printer is free”.
3. Require the agent to test its own output. There are 416 automated tests. On every run, a bot plays all ten levels without a screen, verifies that each one can be cleared, and checks the scoring limit above.
4. Give the agent durable context. A fresh agent session knows nothing about yesterday, so a short file called AGENT_CONTEXT.md holds decisions, and every session reads it first. That is the difference between a collaborator and a very clever goldfish.

The loop even produced audio. Every click, bell, and motor hum is synthesised by a Python script, with not one recorded sample. That output was described as impossible to do alone in twelve days.

What this means for an AIApplications team

The workflow is not gaming-specific. It maps to common business use cases: a product brief becomes the prompt, a clickable prototype becomes the playable output, user testing becomes the complaint cycle, and a project context document becomes the agent memory. The discipline is the product: tests plus context plus feedback.

Common misconceptions to avoid

• Misconception: “You need the most expensive model to build something real.” The case argues the opposite. Maker Lab has 3D graphics, physics, audio, and a test suite, making it a hard application, but the method also works on a Raspberry Pi on a desk.
• Misconception: “Local models are only toys.” With Ollama or llama.cpp, a Raspberry Pi 5 with 8GB or 16GB of RAM can run a small open-weight model locally. It is slower and less capable than cloud AI, but it is private, account-free, subscription-free, prompt-fee-free, and can work with Wi-Fi switched off.
• Misconception: “AI work has to look like a full enterprise system.” A classroom, code club, or kitchen table can use the same pattern. Ask a small model for a Pygame game where a cat catches falling fish, then run it and ask it to make the fish faster every ten points.

For heavier AI tasks, Raspberry Pi also offers AI HATs. Vision work such as spotting people and objects in a camera feed can use the AI HAT+. The AI HAT+ 2 is aimed at projects that need a local language model: its Hailo-10H accelerator delivers 40 TOPS of inferencing performance, and with 8GB of dedicated memory it can run small language models, including Qwen2.5-Coder, while reducing load on the Raspberry Pi itself.

Why the game is free and what that says about trust

Maker Lab is free, with no adverts and no in-app purchases. one button in the settings opens the developer’s JustGiving page for Alzheimer’s Society. That transparency is an application lesson: when AI lowers build cost, distribution can be simple, low-friction, and tied to a clear social purpose.

Maker Lab is raising money for Alzheimer’s Society

The game also asks for working memory, planning, and attention. A player must hold three orders in mind, move across the room, route between machines, and notice a jam. These are the same mental faculties that Alzheimer’s affects. A phone game is not a cure, but if the player enjoys it, giving a few pounds to people working on dementia is presented as a reasonable trade.

The project was described as personal: a path from garage tinkering to Maker in Chief at Raspberry Pi, with one cleaned corner at a time, and a family connection through the developer’s mother, who has Alzheimer’s.

Action advice: copy the workflow on a small scale

Start small: One screen, one mechanic, one afternoon. The advice is deliberately operational.

• Describe the feel and workflow before describing the implementation.
• Test on the real device or environment after every change.
• Record learned decisions in a context file that every session reads first.
• Treat errors as conversation turns, not one-off bug reports.
• Keep the loop human: have fun, build something, and give something back.

The game is out now on the App Store. The builders are encouraged to share what they make, whether on a Raspberry Pi or another device, and to show it at a Maker Faire if possible.

Now the work still has a human tail: level eight still needs play-testing.