> ## Documentation Index
> Fetch the complete documentation index at: https://evalgate.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Run a Playground

> Compare prompt variants side by side on real cases.

# Run a Playground

The Playground lets you try prompt or model changes against real cases before
they touch production. It is the safe place to answer "would this change be
better?"

## Step 1 — Open a Playground

Open `/playgrounds`, then pick the evaluation you want to experiment on. The
Playground opens at `/evaluations/{id}/playground`.

<img src="https://mintcdn.com/evalgate/c7l4ztAi8W3wgmvV/help/images/run-playground-list.png?fit=max&auto=format&n=c7l4ztAi8W3wgmvV&q=85&s=5a4263a0301e3d37244a46c5e3312370" alt="A screenshot of the Playgrounds list with one evaluation selected." width="1440" height="900" data-path="help/images/run-playground-list.png" />

## Step 2 — Add a baseline variant

A Playground needs at least one **baseline** variant — the current production
prompt — to compare against. Add it first.

## Step 3 — Add a candidate variant

Create a new variant with your proposed change. Give it a label like
`shorter-system-prompt`. You can have several candidates at once.

<img src="https://mintcdn.com/evalgate/c7l4ztAi8W3wgmvV/help/images/run-playground-variants.png?fit=max&auto=format&n=c7l4ztAi8W3wgmvV&q=85&s=5ebe81b41cbc0a866a7ac1b7d5e1aca5" alt="A screenshot of the Playground with a baseline and one candidate variant." width="1440" height="900" data-path="help/images/run-playground-variants.png" />

## Step 4 — Run the cases

Pick the cases to run on (a subset is faster), pick the scorer, and click **Run**.
Each row records the variant, case, scorer output, model-call reference, cost,
and terminal status.

<Note>
  A cancelled or partial run is **not** a verified winner. Re-run it before
  trusting the comparison.
</Note>

## Step 5 — Read the comparison

The comparison view shows each variant's pass rate, average score, and cost
against the baseline. Look for a candidate that is better on the metric you care
about without regressing the others.

## Troubleshooting

* **`provider_unhealthy` or `budget_exhausted`** — check the
  [Model Gateway](../evalgate-101/connect-provider-key).
* **All variants score the same** — your scorer may not be sensitive to the
  change. Try a different scorer or add cases that expose the difference.
* **A run is stuck** — cancel it. Cancellation is terminal; do not report a
  cancelled run as success.

## Up next

[Pick a winner](./pick-winner) and turn it into the new baseline.