Design an AI Prompt Playground

Asked by: Anthropic

"Design an AI Prompt Playground" (an Anthropic question) is a trap if you take it at face value. It looks like a CRUD web app — type a prompt, hit run, see output. A weak candidate designs a form and a database and finishes in ten minutes. The interviewer is watching whether you find the three things that are actually hard underneath.

The crux (spend ~60% of your time here). The web tier is table-stakes. The real problems are: (1) multi-model fan-out — run one prompt against several models concurrently and stream each output independently, with partial-failure isolation; (2) reproducibility — a saved run must be re-creatable exactly, which means versioning the prompt and pinning the model version, params, and seed; and (3) cost governance — every run burns GPU tokens, and a playground invites unbounded experimentation, so quotas and caching are first-class, not afterthoughts. The actual inference is a dependency (the GPU serving system), not something you build here.