1 min lesson
What you'd actually build (the sample projects)
Work through the cases in "What you'd actually build (the sample projects)", pairing each signal with the move that fits.
Step 1 of 2
The JD names three example projects and they are not random. Each one maps to a skill the loop tests and to a stage where you'll have to demonstrate it. Learn the mapping and you stop guessing what the interview is really measuring.
New interfaces for reviewing PRs of AI-generated code.
Sprint to build a new product vertical from scratch, e.g. AI bug detection.
Run experiments and A/B tests on millions of users.
The PR-review example is the most revealing of the three. Reviewing a diff a human wrote is a solved interface. Reviewing a diff an agent wrote across nine files, where you trust none of it by default, is an unsolved one. That gap is the kind of invention the role exists to do.
Learn more
Advanced table
Every named project is a proxy for a skill the loop is grading
- JD example
- New UX for reviewing AI-generated PRs
- Skill it really tests
- Front-end craft in TypeScript plus taste for trust-and-verify interfaces
- Loop stage where it shows up
- The 8-hour build onsite, where you design and ship a real feature
- JD example
- Two-week vertical from scratch (AI bug detection)
- Skill it really tests
- Scoping a vague brief and shipping 0-to-1 fast, alone
- Loop stage where it shows up
- The vague-prompt onsite and the 'what would you build in a vacuum' conversation
- JD example
- A/B tests on millions of users
- Skill it really tests
- Eval design, metric choice and reasoning about agent quality at scale
- Loop stage where it shows up
- Product / eval discussion in screens and the onsite presentation
| JD example | Skill it really tests | Loop stage where it shows up |
|---|---|---|
| New UX for reviewing AI-generated PRs | Front-end craft in TypeScript plus taste for trust-and-verify interfaces | The 8-hour build onsite, where you design and ship a real feature |
| Two-week vertical from scratch (AI bug detection) | Scoping a vague brief and shipping 0-to-1 fast, alone | The vague-prompt onsite and the 'what would you build in a vacuum' conversation |
| A/B tests on millions of users | Eval design, metric choice and reasoning about agent quality at scale | Product / eval discussion in screens and the onsite presentation |
Every named project is a proxy for a skill the loop is grading.
Notice what's absent from these examples: nobody is shipping a CRUD endpoint to a spec. The work starts ambiguous and ends in something users touch.
Learn more
Optional practice