This website had reached the point where small changes were taking too much effort. The older theme was aimed towards a different use case, and was starting to look old and dated, so the better option was to replace it with a small in-repo Hugo theme, simplify the layouts, fix the jokes pages, and align the build process with a newer Hugo version.
GenAI was genuinely useful during that work. It sped up repetitive tasks such as rewriting templates, adjusting CSS, and cleaning up old theme behaviour across multiple files.
Recent access to stronger coding models through GitHub Copilot premium has made this kind of work more practical. Copilot Auto mode using GPT-5.4, and Claude Sonnet 4.6 are both effective for coding changes when working from a real repository with clear instructions, and at a much lower token cost than the more powerful Claude Opus 4.6. Across the full overhaul, the overall token cost came to $1.36 and took less than an hour. In practice that means being explicit about the goal, preserving the parts of the site that should stay stable, and validating the result after each meaningful edit.
That said, careful review is the part that matters most. Real problems only showed up during testing: template mistakes, pagination edge cases, stale assumptions from the old theme, and runtime behaviour that no longer matched the data being served. The most useful lesson from the process was that GenAI works well as an accelerator, but only when the work is thoroughly checked and the instructions are strong enough to keep the changes on track.