Anthropic Workbench: First-Use Guide for Non-Coders (2026)

AI Assistant Summary: The Anthropic Workbench is a free browser-based playground for testing prompts against any Claude model (Opus 4.7, Sonnet 4.6, Haiku 4.5). It lives inside the Anthropic Console at console.claude.com. Sign up free, get a small starter API credit, and you can test prompts, compare model outputs, set system prompts, adjust temperature, enable tool use, and save prompt templates — all without writing a line of code. The Workbench is the right tool for non-coders who want more control than Claude.ai offers, prompt engineers iterating on production prompts, and anyone building a Claude integration who needs to see the raw API behavior. This guide walks through the entire first-use flow.

Most people who use Claude do it through Claude.ai — the consumer chat interface. The Workbench is the developer-facing alternative: same models, more controls, no chat history. It’s where prompt engineers iterate, where teams test before shipping to production, and where curious non-coders go when they want to see what’s actually happening behind the conversational interface. The Workbench is free to access; you pay only for the tokens you actually consume in tests.

What you need before starting

• An email address (any account, no work email required)
• ~5 minutes for signup and email verification
• A web browser (Chrome, Safari, Firefox, Edge all work)
• No coding skill required. Workbench is point-and-click.
• No credit card required to sign up — Anthropic provides starter credits.

Step 1: Create an Anthropic Console account

Go to console.claude.com. You’ll see the Anthropic Console signup screen. Sign up with email or Google. After verifying your email, you’re dropped into the Console dashboard.

The first time you sign in, Anthropic shows you a quick onboarding flow asking a few questions about what you want to build. Answer honestly — it doesn’t lock you into anything, but Anthropic uses the answers to tailor the in-app suggestions.

Once you’re in, look at the left sidebar. You’ll see:

• Workbench — where we’re going to spend the rest of this guide.
• Documentation — link to platform.claude.com/docs (our documentation guide curates the most useful pages).
• Settings — account, billing, API keys, organizations.
• Usage — how much credit you have left, current month’s spend.

Step 2: Open the Workbench

Click Workbench in the left sidebar. You’ll see a three-pane layout:

• Left pane: System Prompt. The instructions Claude follows for every message in this session. Empty by default; you can add anything here.
• Middle pane: User Message. Where you type the actual prompt you want Claude to answer.
• Right pane: Settings. Model selector, temperature, max output tokens, tool use toggles, advanced options.

At the top of the Settings pane is the model selector. The default is usually Sonnet 4.6 (the “balanced default” model). Click the dropdown to switch to Opus 4.7 or Haiku 4.5.

Step 3: Your first prompt

Type something into the User Message pane — anything you’d ask Claude. Examples that work well for first runs:

• “Explain the difference between machine learning and AI in plain English.”
• “Write a 200-word product description for a coffee subscription service targeting busy parents.”
• “Summarize this article in 3 bullet points: [paste article].”

Click Run (or hit Cmd/Ctrl+Enter). Claude streams its response into the same pane below your message. The actual cost of the request appears in the right sidebar — usually a few cents.

Step 4: Add a system prompt

This is where the Workbench gets interesting and Claude.ai doesn’t. The System Prompt is instructions that apply to every message in the session. Common uses:

• Role definition: “You are an experienced copywriter for SaaS products. Your responses should be punchy, specific, and never include marketing fluff like ‘unlock the power of’.”
• Output format: “Respond only in valid JSON with the schema: {title: string, summary: string, tags: string[]}.”
• Tone constraints: “Always respond in formal British English. Never use contractions.”
• Context anchoring: “The current date is May 2026. Treat any claim that contradicts this as outdated.”
• Persona: “You are a patient tutor explaining concepts to a high school student. Use concrete examples and avoid jargon.”

Try this exercise: paste a 100-word product description into the system prompt as “You are a copyeditor. Your job is to make text sharper, shorter, and clearer.” Then put a draft paragraph into the User Message. Run. Claude responds as a copyeditor, not as a generic assistant. Change the system prompt to “You are a marketing manager who needs to expand this draft into a longer, more persuasive version” and run the same User Message again. Same Claude, completely different output.

Step 5: Compare models side by side

The Workbench supports running the same prompt across multiple models simultaneously. Look for the “Compare” or “Add Model” button (UI varies slightly with releases). When you add a second model:

• Both Opus 4.7 and Sonnet 4.6 run the same prompt.
• You see both responses side by side.
• You see both costs.
• You can judge whether Opus’s higher price is worth it for your specific task.

This is the killer Workbench feature. You can’t do this in Claude.ai. It’s how prompt engineers decide which model tier to use in production. Our model comparison post covers the general decision tree; the Workbench lets you test it against your actual use case.

Step 6: Adjust temperature and other settings

The right sidebar exposes parameters that Claude.ai hides. The main ones:

• Temperature (0.0 to 1.0): How creative Claude is. 0.0 = most predictable, deterministic, factual. 1.0 = most creative, variable, sometimes surprising. For analysis or classification, set 0.0. For brainstorming or creative writing, try 0.7-1.0.
• Max Output Tokens: Hard limit on response length. Default is high (Sonnet 4.6’s max is 64K). Lower this if you specifically want shorter responses. Higher caps cost more if Claude uses them.
• Stop Sequences: Strings that, if generated, immediately end the response. Useful for structured outputs.
• Top P (sampling): An alternative to temperature. Mostly leave at default unless you know what you’re doing.
• Extended Thinking toggle (Sonnet/Haiku): Turn on if you want the model to reason step-by-step before answering. Costs more but produces better results on hard reasoning. Our extended thinking glossary entry goes deeper.

Step 7: Try tool use

The Workbench supports tool use — giving Claude access to external functions. This is more advanced but the Workbench makes it accessible without coding.

• Click the “Add Tool” button in the right sidebar.
• Define a tool by giving it a name, description, and JSON schema for its parameters. (Workbench has templates for common tools like web search, calculator, code execution.)
• Run a prompt that would benefit from the tool. Claude either answers directly or requests to call the tool.
• The Workbench shows the tool call payload Claude generates.
• You manually provide the tool result, and Claude continues with the new information.

This is also how you simulate building an agent in the Workbench before writing actual code. The official tool use documentation has working examples.

Step 8: Save your work as a prompt template

The Workbench includes a Saved Prompts library. When you have a system prompt that works well, click “Save Prompt” (button location varies). You can name it, tag it, and re-load it for future sessions. Your saved prompts are searchable and survive across sessions.

This is also where you collaborate with teammates: saved prompts can be shared across users in the same Anthropic organization. Useful when you’ve built a great copyeditor prompt and want to make sure everyone on your team uses the same one.

Step 9: Generate code from your prompt

When your prompt is working in the Workbench, click “Get Code” to see the equivalent API call in Python, TypeScript, or curl. The generated code includes:

• The exact model ID you used (e.g., claude-sonnet-4-6).
• Your system prompt and user message.
• All settings (temperature, max tokens, tool definitions).
• An API key placeholder you’ll need to fill in.

Copy this code into your project. It’s the cleanest path from a working Workbench prompt to a working production integration. Our Claude API beginners guide walks through the rest of the deployment.

Workbench vs Claude.ai: when to use which

Use case	Use Claude.ai	Use Workbench
Casual chat / Q&A	Yes	No (overkill)
Daily writing and editing	Yes	No (overkill)
Custom system prompt / persona	Limited	Yes
Compare two models on the same task	No	Yes
Adjust temperature for creativity	No	Yes
Test tool use / function calling	No	Yes
Generate code to integrate later	No	Yes
Share saved prompts across team	Limited	Yes
Long-form ongoing chat history	Yes	No
Files and Artifacts	Yes	Limited

The Workbench is not a replacement for Claude.ai. It’s a complement. Most users will use Claude.ai for daily work and switch to Workbench when they need to nail down a specific prompt or test a model decision.

How much does the Workbench cost?

Access to the Workbench itself is free. You pay only for the tokens you consume in test prompts. Anthropic provides a small starter credit when you sign up, and you’ll see your current balance in the Console’s Usage page.

Concrete cost examples (using current API pricing):

• One typical Workbench prompt on Sonnet 4.6: ~$0.01-0.05 (depending on length)
• One prompt on Opus 4.7: ~$0.05-0.20
• One prompt on Haiku 4.5: ~$0.005-0.02
• An hour of intensive prompt iteration (50-100 test prompts): typically $1-5 on Sonnet, $5-20 on Opus, $0.50-2 on Haiku

The free starter credits cover a few hours of typical use. If you’re iterating heavily, expect to add a payment method or move to a paid Claude.ai plan within the first week or two.

Common first-time mistakes

• Confusing Workbench with Claude.ai. They share models but the interfaces and use cases are different. Workbench is for testing; Claude.ai is for conversational use.
• Forgetting to set the system prompt. Without one, Workbench prompts behave like Claude.ai. Most of Workbench’s value comes from carefully crafted system prompts.
• Cranking up temperature for serious tasks. Higher temperature is for creative work. For accuracy and consistency, default temperature or 0.0 is right.
• Burning through credits on Opus 4.7 when Sonnet would do. Test on Sonnet first; only escalate to Opus when Sonnet’s output isn’t good enough.
• Not saving the prompts that work. The Save Prompt feature exists for a reason. Use it.
• Treating chat history like Claude.ai. Workbench sessions don’t persist by default. Save anything you want to keep.

Frequently asked questions

Do I need a paid Claude.ai plan to use the Workbench?

No. Workbench access is separate from Claude.ai plans. You sign up to the Anthropic Console for free, get starter credits, and pay only for tokens consumed. Your Claude.ai Free/Pro/Max plan doesn’t affect Workbench access.

Is the Workbench safe for sensitive data?

Same data handling as the Claude API. Anthropic doesn’t train on customer API data. If you have specific compliance requirements (HIPAA, GDPR, SOC 2), check Anthropic’s usage policy and consider their Enterprise plan for additional controls. For experimentation with public/sample data, Workbench is fine.

Can I use the Workbench on mobile?

Yes, but it’s not optimized for small screens. The three-pane layout works in mobile browsers, but you’ll have a much better experience on desktop or tablet. For mobile chat use, stick with the Claude.ai app.

What happens to my prompts if I downgrade or stop using the Console?

Saved prompts stay attached to your account. If you stop using the Console, they remain accessible if you log back in. Anthropic only deletes account data on explicit account deletion or per their retention policy.

Does the Workbench support image input?

Yes. All current Claude models support vision (image input). In the User Message pane, click the attach button to add an image, then write your prompt. The model will see the image and respond about it. See the Vision documentation for supported formats.

Can I export my Workbench session?

The “Get Code” feature exports the API equivalent of your session. For chat history, you can copy/paste the messages, or use the Save Prompt feature to capture the system prompt and parameters. There’s no built-in “export full session as PDF” option.

Can my whole team share a Workbench account?

Each user should have their own Console account, but you can create an Anthropic Organization to share API keys, saved prompts, and billing. The Organization features are under Settings → Organizations.

Get the weekly Anthropic feature update

You may also like

• Anthropic Academy: Free Courses to Learn Claude
• Anthropic Documentation: 12 Pages to Bookmark
• Anthropic’s Free Tier: Everything You Can Do Without Paying
• Claude Opus vs Sonnet vs Haiku: Which to Use in 2026
• Claude API for Beginners: Build Your First AI App
• How to Use Claude AI: Complete Beginner’s Guide
• Claude Projects: Organize Your AI Workspace
• Claude Max vs Pro vs Free: Plan Comparison

Sources

• Anthropic. “Console.” console.claude.com (verified May 14, 2026)
• Anthropic. “Workbench documentation.” Part of platform.claude.com/docs
• Anthropic. “Models overview.” platform.claude.com/docs/about-claude/models/overview
• Anthropic. “Tool use.” platform.claude.com/docs/build-with-claude/tool-use
• Anthropic. “Pricing.” claude.com/pricing