The premise started as a workplace joke: photograph someone and have an AI generate a fictional corporate manual about them. “Subject encountered in break room. Optimal operation: compliment the coffee. Known failure mode: discussing weekend plans before 10am.”
That generator ships as one of three built-in makers in MakerMaker, an iOS entertainment app from Niixo. The others produce a mock spec sheet and a fabricated news headline. All three take a photo as input and run it through Gemini 2.5 Flash on Firebase Functions (asia-northeast1). Photos never hit the server — they’re reduced to prompts on-device. iOS 17+ required.
The recursive layer
What became more interesting in practice: a mode where the AI invents the generator concept before generating content.
The button synthesizes a generator from combinations of subject, genre, tone, format, and color palette — roughly 95 million permutations total. Each press produces a generator that didn’t exist before, then immediately runs it.
This is where the product gets both interesting and inconsistent. About 30% of outputs are flat — grammatically fine, tonally dead. When it works, it goes somewhere genuinely strange. The 70/30 ratio is stable enough to ship but worth disclosing upfront.
What we traded off
The UI is Japanese-only. The humor is rooted in Japanese workplace conventions, and we wanted to validate it there before expanding.
UGC moderation runs at the prompt level plus a user-report flow — no automated classifier yet. This is the part we’re most cautious about.
Image storage is zero. Photos of real people shouldn’t leave the device. They don’t.
Why publish the failure rate?
Entertainment apps live on repeat sessions. If users hit flat outputs and assume the product is flat, they churn. Framing the variability honestly — and building the experience around it — is better for retention than overpromising.
Free to download. Point packs from ¥100, monthly pass ¥480.
답글 남기기