Chrome's best advertisement so far

"See, it's not working"

If you're someone who's tried different prompts and still can't get the LLM to fix that one issue in your codebase... that was me before I fixed a performance issue. But one Chrome ad saved me that day.

Coming to the story. Recently I had to pick up a performance issue at work while juggling one of my ongoing tasks that also needed equal attention to meet my timeline (dignity issue, ah 😄). Though hitting the AI_LAZY syndrome pretty hard, I opened a new Claude session and started thinking about where to start. One thing we knew: there was a case involving a large dataset that caused some components in the frontend to create this "Windows screen saver" effect. So I went with that scenario and began tinkering with Claude. As expected, the changes didn't work.

My lazy brain had a plan

As a next step, I went to Chrome and opened the developer tools to analyze the app performance. That's when I noticed the Chrome DevTools MCP ad. So my lazy little brain pushed me to try this.

First step was obviously installing the MCP:

Stealclaude mcp add chrome-devtools npx chrome-devtools-mcp@latest

And I asked Claude to use it to analyze the performance issue (yes, you have to paste the URL of the page. I haven't built a "Shah-Claude MCP" yet 😄). Here's the refined prompt after a few iterations:

StealThe app at localhost:5173 feels janky.
Use Chrome DevTools MCP to:
1. Run a Lighthouse audit and screenshot the score
2. Start a performance trace — then wait for me
3. I'll click around in the app, then tell you when I'm done
4. Stop the trace, analyze it, tell me the top issues and show me a plan before proceeding
5. Apply the fixes, then run a final Lighthouse audit and screenshot to show the improvement

The MCP started doing its thing and then halted, prompting me to click Continue once I'd completed the action that caused the lag. It took a few minutes. It recorded the trace, analyzed the code, and came back with the critical issues and a plan to fix them.

The expected and the unexpected

Fast forward (Reading the plan. the boring stuff). After applying the changes, Claude tested the full flow using MCP and reported that the performance metrics improved from X to Y. Bleh bleh bleh.

Honestly expected to spend another hour fixing whatever Claude broke. But I tested it manually anyway.

Voilà! It worked like butter. No lag, no freezing, nothing. Still, I didn't fully believe it yet. You know.

Doubter to believer

So I started experimenting again. I used Chrome MCP to trigger different actions and also ran my test suites (Playwright) through Claude. I dev-tested everything to ensure that Claude didn't cause any war in my code.

Guess what? It was completely functional and I didn't notice any performance drop. Claude's fixes were textbook, exactly the kind of things you'd normally spend hours tracing with cups of coffee.

THE END.