I was talking to my partner the other day. As the General Manager of a major local restaurant and movie theater, she mentioned they used to feature a rotating beer menu. In the past, they subscribed to a website that provided beer reviews, and as a perk, the site allowed them to easily generate a five-beer “drop menu.”
But when the company canceled the subscription, the rotating menu largely stopped. She and her managers simply didn’t have the time to manually research and build a new menu every time a keg kicked. It just wasn’t a priority.
As I’ve been expanding my abilities around AI as a tool, a lightbulb went off: “This is a job for AI!”
Taking a time-consuming process, researching beers, pulling facts like names, breweries, descriptions, and ABVs, loading them into a template, and producing a printable PDF, is exactly where AI should shine. We constantly read stories about AI having unintended consequences when left without human oversight, and I completely agree: AI is only as good as the people running it. But for something as straightforward as a beer menu? This is AI’s wheelhouse.
Here is the story of how I built an automated solution, the roadblocks I hit, and the lessons you can apply to your own business.
The Prompting Carousel: Claude, Code, and Confusion
I regularly use several AI models—Gemini, Claude, and ChatGPT—and each has its pros and cons. Knowing that the initial prompt is everything, I sat down, wrote the prompt, rewrote it, and edited it again.
Having worked for years in product/program management, IT, and operations, I like having all my requirements upfront. I hate when user stories are constantly edited mid-sprint, so I applied that same discipline here. I’ve found that making change after change just makes the AI model forget what it’s doing and produce worse work.
I decided this was a perfect task for Claude.
I started a chat, and it did a phenomenal job… for about 90% of the build. The initial output worked well, but when I clicked the “print” button on the generated file, it broke.
I went back to the chat and explained the issue. Claude responded with the classic AI pivot: “Oh, I get it, this is what is wrong…” It edited its code and told me to test again.
I did. And I got a completely different error.
I brought that back to the chat. Claude replied: “This time I really got it!” We repeated this loop over and over until Claude finally suggested I move over to Claude Code.
So, I did. But this triggered an entirely new loop. What Claude Code said it needed and what Claude Chat said it needed were actually competing frameworks. I created API tokens, connected to GitHub, and refined the original prompt, but I was still striking out.
The Breakthrough: Switching Engines
Admitting a temporary defeat, I switched over to Gemini. Even with access to advanced models through a Google AI Pro subscription, you still have to actively steer the ship. I knew I needed to use Gemini Canvas to see the code and outputs dynamically.
I pasted the exact same initial prompt I had used in my very first Claude attempt.
Gemini got to 90% in record time… and then it, too, got stuck on the export-to-print function.
I literally felt the air leave my body. A deep exhale. “Here we go again.”
But this time, I tried a different tactic. Instead of explaining the code error, I took a screenshot of the visual issue and dropped it directly into the chat. Gemini instantly figured it out.
After a few small tweaks for fonts and background colors, I had a fully working, automated rotating beer menu. I shared it with my partner. It took a couple more minor edits, but now she and her managers have a tool they can easily use and update on the fly.
I uploaded the code to GitHub Gist if you want to see it: https://gist.github.com/trevshand/ba4395ff20108a46299fad00f0f01954