Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Define "usage" precisely (images created per weekly active user, or weekly creators), find where the funnel leaks, and choose bets that need little engineering and attack the biggest leak. For most image features the leaks are the blank prompt box and the lack of a reason to come back, so starting points, editing your own photos, and remix-and-share loops are the usual winners. Check that cost per image keeps doubling usage affordable.
How to approach it
Ask which metric "usage" means, because doubling images per user and doubling the number of creators are different problems. Ask for the funnel: how many users see the feature, try it, succeed on the first attempt, and come back. With three engineers, every proposal must be small, so the diagnosis decides everything.
A strong answer
Metric. I would target weekly creators (users who generate at least one image they keep, download or share). Images per user can be inflated by people regenerating bad results, which is a sign of failure rather than value.
Funnel, illustratively:
| Stage | Share of weekly active users |
|---|---|
| See the feature | 60% |
| Try it at least once | 15% |
| Keep, download or share a result | 8% |
| Return the following week | 3% |
Two leaks stand out. Of users who see it, three quarters never try. And of those who try, about half never get a result they keep.
Why people do not try. An empty prompt box asks users to be creative on demand, in a skill they do not have. Most people do not know what to type.
Why first attempts fail. Vague prompts produce generic images. Users who type "a birthday card" get something bland and leave.
The blank-box diagnosis is not a guess. When Google added ready-made templates to Google Photos in November 2025, it gave exactly this reason: "Coming up with your own prompt can be daunting", and launched the templates first in the US and India (Google blog, company-reported). The appetite is there once the barrier drops: Google said its Gemini image feature had generated well over 500 million images globally within weeks of launch, with India among the top countries (Google India, September 2025, company-reported).
Bets that fit three engineers.
- Starting points instead of a blank box. A row of tappable templates tied to real moments: birthday cards, festival greetings for Diwali or Eid or Lunar New Year, profile pictures, invitations, memes in the user's language. Each template is a well-written prompt with one or two blanks. Mostly content and front-end work. Attacks the try rate and the first-success rate together.
- Start from your own photo. "Turn this photo into a sticker / a cartoon / a festive card." Editing an existing image gives users an obvious reason and a personal result. If the image model supports editing, this is a few weeks of work.
- Remix and share loops. Every shared image carries a "make your own" link that opens the same template. Sharing brings new users directly into a working starting point, which helps both creators and new-user activation.
I would run 1 first (fastest, biggest leak), then 2, then 3, with an experiment on each.
Cost check. Doubling creators roughly doubles image generation cost, so put a number on it. Google lists its Gemini 3.1 Flash Image model at about $0.0336 per image, and about half that through its batch API (pricing, checked November 2026). Illustratively, 1 million weekly creators making 5 images each is 5 million images, about $168,000 a week at that rate, before any regenerations. If half of all generations are regenerations of results the user did not keep, the cost per kept image is twice the list price. Two mitigations: templates tend to need fewer regenerations than free-form prompts, which lowers cost per kept image, and a daily free allowance with more for subscribers protects margins.
Success criteria. Weekly creators doubled within two quarters, with first-attempt keep rate up and regenerations per kept image down. Guardrails: cost per kept image, and safety reports (more users generating images of real people raises misuse risk, so edits of photos of others need the policy and filters already in place).
What interviewers probe next
"Leadership suggests a better image model instead." Quality helps the 8% who already create. The leak is at "try", where model quality is invisible. Fix the entry point first.
"How do templates work across markets?" Localise by calendar and language: festival templates by country, text rendered in local scripts, which is also a quality test for the model.
How it's scored
| Signal | Weak | Hire | Strong Hire |
|---|---|---|---|
| Metric | Images generated | Active creators | Defines creators by kept or shared output and avoids regeneration inflation |
| Diagnosis | Lists ideas | Funnel thinking | Locates the biggest leaks with numbers and explains the blank-prompt problem |
| Fit to constraint | Big model or platform work | Prioritised ideas | Two or three small bets sequenced by leak size, with cost per kept image and safety guardrails |
Common mistakes
Proposing a new model or a large platform rebuild with three engineers.
Counting regenerations as engagement.
Ignoring that more usage means proportionally more compute cost.