OpenAI Playground is a browser workspace for experimenting with OpenAI API models before integrating them into an application. You can try a request, change the instructions that guide the model and compare the text it produces. This exercise teaches you how to separate a one-time user request from instructions you want the model to follow repeatedly.
You will create two versions of a short customer email using the same model and request. For ordinary conversations rather than API experiments, start with our ChatGPT beginner tutorial.
Before you start
Sign in to the OpenAI API platform and select a project with access to a text model. API billing is separate from ChatGPT subscriptions. Check your available credit or billing arrangement and the API pricing table before generating text. Input and output tokens contribute to the cost, so choose an economical model and keep the experiment short.
Set aside about 10–15 minutes after account setup. You do not need to write an API request for this browser exercise. Use fictional customer information instead of confidential records, and leave external tools disabled.
How to use OpenAI Playground to test a prompt
Step 1: Open Playground
Visit the Playground from the API platform. Open the text prompt workspace. The prompt management guide describes creating a new prompt through the Prompts area. For this exercise, work with a draft; you do not need to publish a reusable prompt ID.
Step 2: Choose your model
Select an available text model in the model configuration. Note its name so you can keep it unchanged throughout the exercise. Retain the default generation settings and avoid adding tools or structured output requirements. The aim is to isolate the effect of your instructions, not to compare different models.
Step 3: Send the baseline request
Enter the following request as the user message. Leave the separate instructions field empty if possible, or retain the same default guidance for both runs.
Write an email telling a customer that their order will ship on Friday instead of Wednesday.
Run the prompt and wait until the response finishes. Copy it to a scratch note. Check whether it preserves both days, whether it adds a reason for the delay that you did not supply, and whether the tone suits a customer email.
Step 4: Add custom instructions
Start a fresh test with the same model and user message, without including the earlier answer in the conversation. In the separate model instructions or developer guidance field, enter:
You write customer service emails. Use plain English and a calm, professional tone. Include a subject line and keep the email body under 70 words. Preserve supplied facts. Do not invent reasons, tracking numbers, discounts or delivery dates. End with a brief apology.
These standing instructions describe how the model should handle the request. The user message still supplies the facts for this particular email. OpenAI's Responses API reference explains the corresponding instructions and input concepts.
Step 5: Run and compare
Submit the same user request again. Read both answers side by side. Check the subject line, word count, shipping days and whether the model introduced an unsupported promise. If the second response misses a requirement, revise the instruction to make that requirement more explicit and try again.
Keep the request and model unchanged during comparison. Differences can also arise from generation variability, so repeat a promising test before relying on the instructions for repeated work.
Example input and expected result
Example of a response that meets the revised instructions:
Subject: Update on your order
Hello,
Your order will ship on Friday instead of Wednesday. Thank you for your patience. We apologize for the delay.
This example preserves the shipping schedule without promising a Friday delivery. It has a subject and a short body, and does not speculate about the cause. Use these properties as your checklist when evaluating your own outputs. A more detailed answer may sound friendly but be less suitable if it introduces unsupported details.
Three tips for better results
- Separate facts from rules. Put the individual order details in the user message and reusable writing rules in the instructions.
- Make constraints measurable. A word limit and a list of facts to preserve are easier to check than a vague request to be better.
- Record your configuration. Save the model, instructions and sample inputs together so you can compare future changes fairly.
Common beginner problems
The request fails because of billing or quota
Check your API project, available balance, usage limits and model access. A ChatGPT subscription alone does not provide API credits. Avoid repeatedly submitting a failed request before resolving the error.
The result ignores a constraint
Remove conflicting instructions and rewrite the constraint as a concrete rule. Test with the original sample again, then try a different customer request to check that the rule generalizes.
The second response resembles the first
The baseline may already satisfy your added rules. Try another realistic input while keeping both configurations fixed, and evaluate whether the revised version consistently preserves facts.
What to try next
First, reuse the instructions with a fictional order cancellation. Second, add a specific formatting constraint and compare the results using our guide to better AI prompts. Third, if you are developing an application, explore the Responses API guide to understand how to pass instructions and input in code. Keep API credentials out of public documents and browser-side code.
You can also compare this approach with testing a prompt in Google AI Studio.
Frequently asked questions
Is OpenAI Playground free?
Model usage is billed through the API platform according to the selected model and your account's billing arrangement. Check your balance and current API pricing before testing.
Do I need a ChatGPT Plus subscription?
No. ChatGPT subscriptions and API platform billing are separate. You need access to the API platform and an eligible model.
What is the difference between instructions and the user message?
Instructions establish the behavior and constraints for the model; the user message supplies the current task and its details.
Do I need to publish a prompt to test it?
No. Experiment with a draft and compare responses before deciding how to reuse the configuration.