Use the playground
Reading a trace tells you what went wrong. The next move is to try a fix, and you want that loop to be fast: change the prompt, see the output, decide. That is what a playground is for.
What a playground is
A playground is a scratchpad for one call. You change the prompt or the model, run it on a single input, and see the output immediately. No code, no deploy. In Respan this is the Playground, where you can also run prompt variants side by side.
The manual loop
Pair it with the trace you just read:
- Read the failing trace and form a hypothesis (“it ignored the instruction to answer in one sentence”).
- Open that input in the playground, make the change, and run it.
- Look at the new output. Better? Keep it. Worse? Try again.
You can go around this loop in seconds, which is exactly why prompt work should start here rather than in code.
Playground vs experiment
The playground has the same blind spot as reading a single trace: it only tells you about the one input in front of you. Your prompt is better on this case, but is it better on average, and did it break anything else?
That is the line between a playground and an experiment:
Prototype in the playground; when you need to trust the result, promote it to an experiment. The same change, measured across many cases instead of one.
What this gives you
- A hypothesis-to-answer loop in seconds for a single case.
- A cheap place to explore before you spend a scored run.
- A clear handoff. The moment “looks better” is not enough, you already know the next step is an experiment.
Next steps
- Turn a promising prompt into a measured result with your first systematic eval, evaluate a single LLM call.