Design Prompt Playground

Design a Prompt Playground (like ChatGPT Playground) for a large language model. The product must let users edit a system prompt plus a user message, click “Run”, and immediately receive a streamed model response. Each run is independent—no conversation memory. Users should be able to tweak model parameters (temperature, top-p, max tokens, model selection) and compare outputs side-by-side. The UI should support split-view panes so the same prompt can be run with different parameter sets simultaneously. All prompt configurations and outputs must be versioned, shareable, and evaluatable: implement git-like versioning for prompts, allow batch execution against test datasets, and surface automatic metrics (token count, estimated cost, latency, response length, keyword presence, sentiment). Support prompt templates with variables (e.g., {{user_name}}). The system must stream tokens to the client in real time via SSE, show live token/cost/latency counters, and allow mid-stream cancellation. Finally, design for global scale: thousands of concurrent users, 100 ms TTFT target, and cost-efficient use of upstream LLM provider rate limits.