Anthropic Console provides a browser-based environment for experimenting with Claude models before using them in an application. Its Playground, formerly called Workbench, lets you try prompts and refine instructions without writing an API integration. The goal of this tutorial is to turn a short set of meeting notes into an actionable handover, then improve the prompt so that it handles missing information correctly.
The Console is intended for developer workflows. If you simply want to chat with Claude or revise a document, start with our Claude beginner guide instead.
Before you start
Sign in through the Claude developer platform. You need access to an organization or workspace authorized to use a Claude model, and sufficient API credits or an appropriate billing arrangement. A subscription to the consumer Claude app does not automatically pay for Console model usage.
Review the official model pricing before running the exercise. Input and output tokens are billed at model-specific rates. Pick an affordable available text model and keep the requests short. Allow about 10–15 minutes for the exercise after account and billing setup. Use the fictional notes below rather than confidential business information.
How to test a Claude prompt in Anthropic Console
Step 1: Open Playground
Open Playground from the Console navigation. Older tutorials may call this area Workbench; Anthropic documents product changes in its platform release notes. Start a new prompt experiment rather than editing an existing production prompt. Select the workspace in which you intend to test.
Step 2: Select a Claude model
Choose a text-capable Claude model available to your workspace. For short extraction and rewriting tasks, begin with an economical option rather than selecting the most expensive model automatically. Leave other generation settings at their defaults. Record the model name so you can repeat the experiment using the same configuration.
Step 3: Enter a baseline prompt
Enter the following request and notes in the user message area:
Turn these meeting notes into a short handover for the project team: Maya will send the revised design on Tuesday. Alex needs to review the budget, but no deadline was agreed. The client presentation is scheduled for Friday. Someone needs to reserve a meeting room, but nobody was assigned.
Run the prompt and wait for the response to finish. Copy the result into a note. Identify whether Claude distinguished assigned responsibilities from unassigned tasks and whether it treated the presentation as an event rather than inventing an owner or deadline.
Step 4: Refine the instructions
Keep the model and original user message unchanged. Start another test without including the first model response. In the system prompt area, add the following guidance:
You prepare project handovers. Return a concise list of actions with three fields: task, owner and deadline. Write "Not assigned" when an owner is missing and "Not agreed" when a deadline is missing. Put scheduled events in a separate section. Preserve the supplied facts and do not infer names or dates.
The system prompt defines recurring behavior; the user message supplies the particular notes. Anthropic's prompt engineering documentation explains how clear instructions and examples improve task definition.
Step 5: Run the refined prompt and compare
Run the revised configuration. Compare the outputs using the same criteria: all four notes represented, correct owners, correct deadlines, and no invented details. Check that Friday appears under scheduled events and is not reused as the budget deadline. A more polished answer is not necessarily more accurate.
If a field is missing, make the requested structure more explicit before another run. Save the prompt text and model name together so you can reproduce the configuration with a different set of notes.
Example input and expected result
A suitable example result for the refined prompt is:
| Task | Owner | Deadline |
|---|---|---|
| Send revised design | Maya | Tuesday |
| Review budget | Alex | Not agreed |
| Reserve meeting room | Not assigned | Not agreed |
Scheduled event: Client presentation — Friday.
This format makes the unresolved decisions visible. The team can assign a room owner and agree on the budget deadline rather than accidentally treating AI-generated guesses as commitments. Compare your result against the supplied notes, not just the appearance of this table.
Three tips for better results
- Separate rules from input. Put reusable formatting and accuracy instructions in the system prompt and changing business facts in the user message.
- Define missing information explicitly. Ask for a consistent placeholder instead of letting the model guess.
- Change one thing at a time. Keep the model and sample input fixed while improving the instructions. See our guide to writing better prompts.
Common beginner problems
You cannot run a prompt
Check the selected organization, workspace permissions, API billing and available model access. Insufficient credit or a rate limit requires a different remedy from an invalid prompt. Resolve the displayed error before submitting repeatedly.
The response invents a deadline
State that dates must come directly from the supplied notes. Include an example of a task with an unknown deadline and the exact placeholder you expect.
The output ends before all tasks appear
Check the configured output token limit if exposed for your model. Shorten unnecessary instructions or allow more output where appropriate. Remember that increasing output length can increase cost.
What to try next
First, test the same prompt with new notes containing deliberately missing owners and dates. Second, try a different model while preserving the successful prompt and compare accuracy and cost. Third, explore the Claude API documentation if you want to automate this handover process. Keep API credentials private.
For broader model selection and usage considerations, read our Claude overview, or compare the same experimentation approach with OpenAI Playground.
Frequently asked questions
Is Anthropic Workbench still available?
The Console prompt-testing workspace is now called Playground. Older Workbench tutorials refer to the previous name.
Is the Console included in Claude Pro?
Consumer Claude subscriptions and developer API usage are separate billing arrangements. Check your Console organization's billing settings before running model requests.
Do I need to write code?
No. The browser Playground is intended for prompt experiments. Building an application with the Claude API is a separate step.
Why use a system prompt?
It provides reusable guidance for how Claude should handle the task, while the user message contains the specific information to process.