GUIDES
Claude Can Actually See 📚
Claude has eyes and almost nobody uses them. The second you start sending it screenshots, it changes what you can build. Here are the two ways I use it every week.
Here is the thing most people miss: Claude can see images. You can drop a screenshot, a photo, or a design right into the chat and it reads it like you would, the layout, the colors, the spacing, the vibe. Most people only ever type words at it, so they are using half the tool.
Once you start feeding it pictures, two superpowers open up, especially if you ever build anything visual like a website, an app, or social posts. Let me show you both, with the exact words to copy.
Where This Works
In the regular Claude app, you just drag an image into the message box or paste it (Cmd/Ctrl + V). On your phone, attach it like a photo. In Claude Code, you can drag an image into the terminal or paste its file path. Either way, then you talk about what you sent.
Design From A Picture, Not A Paragraph
When you are building something visual, stop trying to describe the look you want in words. "Make it modern and clean and premium" means nothing, every AI builds a different bland version of that. Instead, show it.
Find a website, app, or post whose design you love. Take a screenshot, drop it into Claude, and say "make it look like this." It pulls the actual design patterns straight from the image, the spacing, the type, the color feel, the layout, so what you get back feels custom and premium instead of like every other AI-generated page on the internet.
My two rules for this: First, pull inspiration from a totally different industry than yours, not a direct competitor. If you sell skincare, screenshot a beautiful architecture studio or a coffee brand, not another skincare site. That is how you end up looking like nobody else instead of a copy of the brand next door. Second, treat what you get as a starting template, then put your own spin on it. The screenshot gets you 80% of the way past the blank page, you make it yours from there.
Design-from-reference prompt
Pro Move
Drop in two or three screenshots at once and say "take the layout from the first, the color feeling from the second, and the typography from the third." Mixing references from different places is the secret to a look that feels original instead of borrowed.
Make Claude Check Its Own Work
This one is the real game-changer, and it is how I build now. When Claude makes something visual, do not just take the first version and hope it looks good. Make it look at its own work the way a person would, and fix what is off, before it ever reaches you.
I tell Claude to build the thing, then take a screenshot of the actual result, then look at that screenshot and judge whether the layout actually looks right, is anything overlapping, off-center, too cramped, ugly. It spots the problems, fixes them, screenshots again, and repeats until it is genuinely good.
So by the time it hands me a "first" version, it has already done three or four rounds of building and checking itself. That version is night-and-day better than what I used to get, because all the obvious stuff got caught and fixed before it ever landed on my screen. You stop being the one who finds every little visual bug.
Self-check loop prompt
The Honest Setup Note
For Claude to screenshot a live website on its own, it needs a tool that can open the page, which in Claude Code usually means a browser tool like the Playwright connector (it can set this up for you, just ask: "add a way for you to open and screenshot the page"). If you do not have that yet, you can run the loop manually: you take the screenshot, paste it back in, and say "here's how it actually looks, what's off and fix it." Same powerful result, you are just the camera.
The Real Win
Claude having eyes turns it from a thing that follows instructions into a thing that can judge its own output. Show it what good looks like, and make it check itself against that. That one shift is the difference between "AI slop" and something you are actually proud to put your name on.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.