The missing piece for Claude, Lovable and Bolt: visual feedback for AI coding

Building a site with AI takes minutes. Fixing the small UI issues takes days. Here's how I point at problems on a live site and let Claude fix them.

The Missing Piece for Claude, Lovable and Bolt video thumbnail by Ali Amin
On this page
  1. Watch the video
  2. Why prompting your way through UI fixes doesn't scale
  3. How visual feedback works
  4. Setting it up
  5. Let Claude read and fix the list
  6. Let clients leave feedback without installing anything
  7. Why this matters for client work
  8. Follow along

Anyone can build a good-looking website in minutes with Claude, Lovable, Bolt or ChatGPT. You type a prompt and the site is ready.

Then the refinement loop starts. Darker button background. Resize the hero image. Fix the border radius. Update the font. Every tiny change means explaining yourself to the AI all over again. What should take minutes takes hours, and if you build for clients, the back-and-forth review cycles can eat your whole week.

The real problem is simple: you can see exactly what needs fixing, but your AI coding agent can't see it with you. In this post I'll show you how I solve that with visual feedback for AI coding: you point at any element on your live site, and the AI gets the exact context it needs to fix it.

Watch the video

Watch: The Missing Piece for Claude, Lovable and Bolt

▶ Watch "The Missing Piece for Claude, Lovable & Bolt" on YouTube

Disclosure: this video is sponsored by Skyelight.

Why prompting your way through UI fixes doesn't scale

I built a site with Claude Code. The first build was strong, but one page alone had six or seven issues:

  • A button color you could barely see
  • Two different fonts on the same page
  • Footer buttons that disappeared into the background
  • A button with the wrong corner radius

Prompting Claude line by line, with screenshots, works for a handful of fixes. But real projects with multiple pages give you dozens, sometimes hundreds, of small UI adjustments. Keeping that list organized becomes a job of its own.

How visual feedback works

Skyelight fills that gap. You click any element on your live site and leave a comment. Behind the scenes, it captures the technical context: which element it is, where it sits, and the code details the AI needs. Each comment is automatically sorted as a bug, feedback or an idea, and you can attach a screenshot too.

You never have to touch that code yourself. It's there so your AI understands the problem without you describing it. That's the whole idea behind visual feedback for AI coding: show the problem instead of explaining it.

Setting it up

  1. Create a free account at skyelight.ai with Google or email.
  2. Create a project and add your website's link.
  3. Install the Chrome extension, open your site, pick your project, and click the + icon.
  4. Hover over any element to highlight it, click it, and write your comment.

Everything you pin shows up as a list in your workspace. On a single page I had six fixes queued in a couple of minutes.

Let Claude read and fix the list

Copying that list into a prompt would defeat the point. Instead, connect Skyelight to Claude through MCP:

  1. In Claude, go to Connectors and add a custom connector named Skyelight, using the MCP address from Skyelight.
  2. Sign in and allow the connection.
  3. Prompt Claude with something like: "Go into my Skyelight project called AI Engineer Academy, pull the open fixes, and update the website."

Claude read every item, fixed the button colors, the corner radius, the footer text and the borders, then marked each item as resolved in Skyelight. Anything still open stays under the open filter, and you can reopen an item if a fix isn't right.

Tip: if Claude keeps asking permission for every action, open Manage connectors, find Skyelight, and switch it from needs approval to always allow. Only do this for tools you trust.

Let clients leave feedback without installing anything

Most clients won't install a Chrome extension, and they shouldn't have to. Here's how to make it effortless for them:

  1. Copy the one-line script from Skyelight's docs.
  2. Ask Claude to add it to your site's code, then redeploy.
  3. Turn on the share toggle in Skyelight and send the link to your client.

Your client opens the link, clicks the Skyelight icon in the corner, enters their email and pins comments as a guest. Their feedback lands in your workspace with their email, the screenshot and the classification. Then you give Claude one instruction: check what my client added and fix it.

Why this matters for client work

The old process looked like this: long text prompts, screenshots, sending the client a link, getting their feedback back as more screenshots in a Google Doc, then translating all of it for the AI. Now it's precise pins on the live site that your AI reads directly.

You can start with Skyelight for free. Connect it to Claude and see how much time it saves on your next build. If you're writing the content for those sites too, read my post on the copywriting mistakes that kill AI websites.

Want to build and sell client sites and AI web apps like this? That's what I teach step by step inside AI Engineer Academy. Here's what's inside the academy.


Follow along

Keep building.

Ali Amin

The community

Learn it with me, not alone.

AI Engineer Academy on Skool is where the courses, live calls and other builders are. Learn at your own pace, with people who are building too.