What We Learned Building Our Own Site With Claude (Not the Generic "AI Changes Everything" Take)

Anthropic's Claude AI assistant assisted in the creation and upkeep of a site. This practical experience provides an answer to a common query: does AI actually work for production tasks or is it merely a demonstration? Below are three concrete lessons gained, rather than the cliché "AI changes everything."

1. AI's effectiveness relies on its ability to verify rather than guess. A mobile layout issue surfaced where content was overflowing on phone screens. Automatically diagnosing the CSS code seemed like the logical solution. However, that approach often overlooks the true problem. The issue was due to a legacy stylesheet containing five conflicting @media breakpoints.

To reliably identify this, live browser inspection at the correct viewport size was necessary, checking computed styles instead of merely reading the source code and assuming.

2. The shortcomings of the tool itself can be as important as its abilities. Producing branded social media graphics seems straightforward - a photo, a logo, and some text. However, doing this programmatically introduces several systems: downloading an image, encoding it, passing it to a script, and finally exporting the result.

An early attempt corrupting an image silently - without any error message, just a slightly incorrect file that worked until you checked the byte count against the original. The fix wasn't an improved AI; it was a different process: moving binary data file-to-file through the shell, and verifying file sizes instead of trusting that "no error" means "correct."

This is an engineering discipline issue, not a model-quality one, and it's the type of thing you only discover through actual implementation and shipping, not by reading about AI coding in theory.

3. True technical complexity is where AI-assisted work truly shines. Developing a branded PowerPoint template required understanding the difference between a .potx template and a regular .pptx file. This distinction isn't about the file extension; it's a single content-type declaration hidden inside the file's internal XML. Getting this wrong causes PowerPoint to treat the template as a normal presentation.

This isn't about creative writing; it's about specific, unassuming technical details that distinguish AI-assisted work that actually ships from work that appears impressive in a demo but fails once real file formats are involved. The honest conclusion is that AI doesn't replace engineering judgment; instead, it accelerates it if the surrounding process is disciplined - verifying instead of assuming, understanding tools' actual failure modes, and not shying away from unglamorous technical details. This standard should be applied to any senior engineer, whether human or AI-assisted.