Concepts

Playground

The playground is a quick loop for iterating on a prompt: write it, pick a model and parameters, run it, read the output, adjust, repeat.

Controls

Model — one of the available models (gpt-4o, gpt-4o-mini, claude-sonnet-5, claude-haiku-4-5, gemini-2.5-pro).
Temperature — 0 to 1, higher values produce more varied phrasing.
Max tokens — an upper bound on the completion length.

Run history and comparison

Every run is added to a history list. Click a past run to pull it into a side-by-side comparison with the latest one — useful for judging whether a prompt tweak or a different model actually changed the output.

Mocked completions

In this preview, the "Run" button calls a mock completion endpoint that returns canned text with simulated latency — no real model API calls or cost are involved. A production deployment would swap this endpoint for a real model provider call.

Playground vs. Evaluations — when to use which

They look similar but answer different questions. Playground is for exploring: one prompt, one output, no ground truth, fast iteration. Evaluations is for verifying: the same prompt template run against an entire golden dataset, scored against known-correct answers. Use Playground to find a prompt that seems to work, then move it into a Prompts version and confirm it holds up across every test case with a real evaluation run — a prompt that looks great on one example can still fail on the edge cases a dataset catches.