Skip to main content
The Playground runs one model call on your own provider key, so you can try a change without touching your agent. Open it from the sidebar. Running needs the editor role.

​
Three sources

Source picks where the prompt comes from. It opens on Scratch, unless a link chose another. Trace and Skill show a Comparison: Original — shipped beside New — your replay. Only that one call re-runs, so the run’s later steps don’t. Trace sends just the system and user messages, while Skill sends the run’s whole prompt. Judge with your key adds an optional opinion on the two.

​
Test a skill edit against a real run

1

Open the skill

On the skill’s page, choose ⋯ More → Test against trace.
2

Edit and pick a run

Edit the Skill body. The test: picker lists recent runs that loaded the skill.
3

Run and compare

Run sends the prompt that run really received, with your edit in place of the skill’s text.
4

Save

Save as new version saves the edit. Agents that follow the latest version get it on their next run. Agents pinned to a version don’t.
If the Playground can’t find where a run received the skill, Run is off and the page says “Can’t run this edit against this trace.” Pick another run, or run the agent again to capture a new one. One replay can’t show that an edit helps. Benchmark draft runs the skill’s eval suite with and without the body, on DecimalAI’s key and counted against your plan. It saves any unsaved edit as a new version first. To compare setups on your own cases, see Test a change.

​
Settings and keys

Run settings has Model, which opens on gpt-5.4-mini, Provider and Temperature (0 to 2, default 0.7). Picking a run fills them from its call. ⌘↵ or Ctrl+Enter runs. Runs and Judge with your key use your own OpenAI, Anthropic or Gemini key. An organization owner or admin adds keys under Settings → General → Provider keys (BYOK).