Set Up Your AI Design Workspace in 20 Minutes

Start here

If you are new to AI design workflows, do not start with MCP, agents, custom skills, or design-token pipelines.

Start by making sure the basic workspace works.

This guide gets you to one simple outcome: you can open a real project folder, ask AI to inspect it, and understand whether the response is useful.

The setup order

Use this order if you are starting from zero.

1. Create or confirm your Claude account

Go to claude.ai and sign in.

If you only want to write prompts and critique screenshots, the chat interface is enough. If you want Claude to read folders, edit files, run commands, and maintain a design system project, you need Claude Code, which requires a paid Claude plan.

Do not spend an hour comparing every model. Pick one tool and run the first exercise.

2. Install Cursor

Download Cursor from cursor.com/downloads. Install it like any other app.

Cursor matters because it gives you a safe place to review AI edits. You can open a file, ask for a change, inspect the diff, and accept or reject before anything becomes final.

For most designers, this is less scary than starting in the terminal.

3. Optional: install Claude Code

If you already have a paid Claude plan, install Claude Code now. Open your terminal and follow Anthropic’s current quickstart instructions.

You do not need to master the terminal. For this curriculum, the main terminal moves are:

cd path/to/project
claude

That means: move into a project folder, then start Claude Code there.

4. Pick your first project folder

Choose something real but low-risk.

Good first folders:

• A copy of your design system repo
• A folder of exported design tokens
• A docs folder with component guidelines
• A markdown folder with research or product notes
• A throwaway copy of a client project

Avoid your most sensitive production repo on day one. You are learning the loop.

The first sanity check

Open your folder in Cursor or Claude Code and ask:

What is in this project? Give me a short map of the important files, the likely purpose of the project, and the first three things I should read before making changes.

A useful answer will mention real files by name and explain why they matter. A weak answer will stay generic: “This appears to be a project with components and configuration files.”

The sanity check tells you whether AI can actually see your context.

When the sanity check fails

A generic answer is not a dead end. It is a signal. Almost every first-day failure traces back to one of three causes, and each has a quick fix.

• The answer names no real files. The tool probably cannot see your folder. In Cursor, confirm the file tree is actually showing in the Explorer, not an empty sidebar. In Claude Code, confirm you ran claude from inside the project folder, not your home directory. Re-open the correct folder and ask again.
• The answer is generic but the files are visible. The model can see the folder but did not look. Push back in the same chat: “Use exact filenames from this folder. If you cannot read the files, say so explicitly instead of guessing.” A real tool will either cite files or admit it is blocked. A vague second answer means the context is not reaching it.
• The folder is too big or too noisy. If you opened a giant repo with node_modules, build output, and thousands of files, the important files get buried. Start with a smaller folder: just the tokens, just the docs, just one component directory. You can widen the scope once the loop works.

If all three checks pass and the answer is still useless, the problem is usually the prompt, not the setup. That is a good problem to have on day zero, because prompting is a skill you can practice immediately.

What not to set up yet

Skip these on day zero:

• MCP servers
• GitHub Actions
• custom Claude skills
• Figma API tokens
• design-token sync pipelines
• browser automation
• private package publishing

Those are powerful later. Today, they create false complexity.

Your only job is to get one working loop: open folder, ask, read, decide.

The minimum viable AI workspace

You are ready to continue when you have:

• One AI chat account you can access
• One editor installed
• One real folder open
• One successful project-map response
• One note where you write down what confused you

That last note matters. Most AI workflow problems are not tool problems. They are hidden assumptions. Write down every place where you get stuck so you can turn it into setup context later.

Exercise

Run your workspace sanity check

20 min

1. Open one real folder in Cursor or Claude Code

   Pick a folder that contains real work but is safe to experiment with. Open it in Cursor, or start Claude Code from that folder. Do not create a blank project just to feel safe. The point is to see how AI behaves with actual context.

   • The file tree is visible in Cursor or available to Claude Code
   • You know whether this is a copy, sandbox, or production folder
   • You can name the top-level files or folders before asking AI about them

2. Ask for a project map and judge the answer

   Ask the sanity-check prompt from this guide. Read the answer. Highlight every sentence that mentions a real file or folder. If the answer does not cite specific project details, ask again with: “Use exact filenames from this folder. If you cannot see files, say so.”

   • The answer names real files from your folder
   • The explanation helps you decide what to read first
   • You know whether the AI has enough context to help with a real task