Talking to Claude Like a Pro

The Number One Mistake

The biggest mistake people make with Claude Code is being too vague. It’s like telling a taxi driver “take me somewhere nice” versus “take me to 123 Main Street.” Both work, but one gets you exactly where you want to go.

Vague prompt:

“Make me a website”

Specific prompt:

“Build me a landing page for a dog walking business called ‘Happy Paws.’ Include a hero section with a tagline, a services section with three pricing tiers, a testimonials section with placeholder reviews, and a contact form. Use a warm, friendly color scheme with rounded corners.”

The vague prompt works — Claude will build something. But the specific prompt builds exactly what you pictured. The secret is to describe what you want the way you’d describe it to a contractor renovating your kitchen: materials, colors, layout, the works.

The Before and After

Here are real examples showing how a small change in wording makes a huge difference:

Before: “Add a login page” After: “Add a login page with email and password fields, a ‘Remember me’ checkbox, a ‘Forgot password?’ link, and a sign-up link at the bottom. Use the same styling as the rest of the site. Show a red error message if the credentials are wrong.”

Before: “Fix the bug” After: “When I click the ‘Add to Cart’ button on the product page, nothing happens. The console shows ‘TypeError: Cannot read property of undefined.’ Find the bug, fix it, and verify the cart updates correctly.”

Before: “Make it look better” After: “The page looks too plain. Add more whitespace between sections, increase the font size of headings, add subtle box shadows to the cards, and use a consistent color palette based on navy blue and white.”

Build a recipe manager app. I want to add recipes with a title, ingredient list, and step-by-step instructions. Include a search bar that filters recipes by name or ingredient. Use a clean, minimal design with a white background and soft shadows on the recipe cards. Save everything to localStorage.

Give Claude a Way to Verify

One of the most powerful techniques is telling Claude to check its own work. Instead of hoping the code works, tell Claude to prove it works.

Write a function that calculates the total price of items in a shopping cart, including tax. Write tests for it — test empty carts, single items, multiple items, and different tax rates. Run the tests and fix any failures.

That last sentence — “Run the tests and fix any failures” — is doing serious heavy lifting. It turns Claude from a code writer into a code writer who debugs their own work. Always add verification criteria when you can:

• “Run the tests and fix any failures”
• “Build it and make sure there are no errors in the console”
• “Start the server and verify the homepage loads correctly”
• “Commit only if all tests pass”

Pro Tip

Think of verification criteria as a checklist you’d hand to an employee. Instead of “paint the room,” you’d say “paint the room blue, let it dry, and check for any spots you missed.” Same energy.

Reference Files with @filename

When your project has multiple files, you can point Claude at a specific one using the @ symbol. This focuses Claude’s attention exactly where you need it:

Look at @src/components/Header.tsx and make the navigation responsive. On mobile, it should collapse into a hamburger menu. Keep the same styling.

You can reference multiple files too:

Compare @package.json and @package-lock.json. Are there any dependency version mismatches? List them and fix any that could cause problems.

Claude already sees your whole project, but @filename is like tapping someone on the shoulder and saying “look at this specifically.” It reduces confusion and speeds up responses.

Plan Mode: Think Before You Build

Sometimes you don’t want Claude to start building immediately. Maybe you want to discuss the approach first, or you’re not sure what you need. That’s what Plan Mode is for.

Press Shift + Tab to toggle Plan Mode on. When it’s active, Claude will think and plan without making any changes to your files. It’s like asking an architect to sketch blueprints before swinging a hammer.

Once you like the plan, turn Plan Mode off (Shift + Tab again) and tell Claude to execute it.

Pro Tip

Use Plan Mode when you’re working on something complex or unfamiliar. It’s the difference between “start building” and “let’s talk about what we’re building first.” Both have their place.

The Pro Workflow: Explore, Plan, Implement, Commit

The best Claude Code users follow a rhythm. It goes like this:

1. Explore — Ask Claude to understand the project. “What does this codebase do? Walk me through the architecture.”
2. Plan — Turn on Plan Mode. “I want to add feature X. How should we approach this?”
3. Implement — Turn off Plan Mode. “Go ahead and build it. Run the tests after.”
4. Commit — “Commit these changes with a descriptive message.”

This loop keeps you in control. You understand what’s happening at every step, and Claude does the actual building. It’s like being the general contractor on a construction site — you make the decisions, Claude does the labor.

Paste Screenshots and Error Messages

Claude Code can read images. If you see an error on screen or a design you want to replicate, just paste a screenshot directly into the chat. Claude will analyze it and respond.

Common use cases:

• Bug screenshots — “Here’s what the page looks like. The button should be aligned to the right, but it’s stuck in the center. Fix it.”
• Design mockups — “Here’s a Figma screenshot. Build this layout.”
• Error messages — Paste the full error. Claude traces it to the root cause.

You can also paste error messages as text. The more context Claude has, the faster it solves the problem.

I'm getting this error when I start the app: 'Module not found: Can't resolve ./components/Header'. I haven't changed any file names. What's wrong and how do I fix it?

Pipe Data Into Claude with the Command Line

For more advanced usage, you can send data directly to Claude using the pipe operator. This lets you feed Claude the contents of files, command output, or anything else:

The -p flag sends a one-off prompt without starting an interactive session. It’s perfect for quick questions when you don’t need a full conversation.

Pro Tip

The pipe trick (cat file | claude -p "question") is great for analyzing log files, CSV data, configuration files, or anything you want Claude to read and interpret quickly.

Try It Yourself