Sunday Coffee & Code: From Google Sheets to a working web app in Google AI Studio

Today’s experiment was taking the Google Sheets/App Script and Google Cloud Run solution I’d already built and seeing how far I could push it into a fuller web app using 𝗚𝗼𝗼𝗴𝗹𝗲 𝗔𝗜 𝗦𝘁𝘂𝗱𝗶𝗼 𝗕𝘂𝗶𝗹𝗱 (and of course learn about 𝗕𝘂𝗶𝗹𝗱 in the process. I started by having Gemini create a PRD from the Clou Run version (with HTML and Vision scan fall back), and used that in AI Studio to generate a more complete app. 𝗧𝗵𝗲𝗻 𝗰𝗮𝗺𝗲 𝘁𝗵𝗲 𝗱𝗲𝗯𝘂𝗴𝗴𝗶𝗻𝗴. Spent time letting the agent try to fix the APIFlash usage itself - I gave up waiting after 3 attempts and gave it a working example - was fine after that. One feature I did like: when running the code in Studio, it caught an error and offered a 𝗙𝗶𝘅 button. That was cool. Next was a Google Gemini API key problem. Interestingly, it responded by adding more logging to expose the underlying error, which was a sensible approach. It looked like secret handling was the issue, so I suggested it work around it by creating a new variable and pasting the key directly. Not as elegant as selecting a project and a secret, but effective. GMail integration needed a few more pieces: extra keys, a Google OAuth 2 Web App, callback URLs, and a new Client ID and Client Secret. The most persistent issue was the app not reliably recognising that I had already connected my Google account. After several attempts, I asked for more logging so I could inspect console messages in Chrome. That led me into watching cookies and session IDs in F12, which felt a bit less 𝘭𝘰𝘸-𝘤𝘰𝘥𝘦/𝘯𝘰-𝘤𝘰𝘥𝘦 than advertised. In the end, It was eventually fixed and the account connection flow worked much like the Google Sheets version. 𝗥𝗲𝘀𝘂𝗹𝘁: a working basic web app for website scanning and email drafting with a reasonable GUI rather than a batch job or App Script (screenshots attached). One other observation: the app initially used Vision to scan the test site. It did reasonably well, but not as accurately as an HTML scan. Since Vision was only meant as a fallback, I switched to HTML scan with Vision fallback - and that worked a treat. All in all, about 3 hours to get to a working app. A few reflections: • 𝗚𝗼𝗼𝗴𝗹𝗲 𝗦𝗵𝗲𝗲𝘁𝘀 + 𝗔𝗽𝗽 𝗦𝗰𝗿𝗶𝗽𝘁 was the easiest, but limited by the 6-minute runtime. • 𝗚𝗼𝗼𝗴𝗹𝗲 𝗖𝗹𝗼𝘂𝗱 𝗥𝘂𝗻 batch job was simple and robust running in the background. • 𝗚𝗼𝗼𝗴𝗹𝗲 𝗔𝗜 𝗦𝘁𝘂𝗱𝗶𝗼 𝗕𝘂𝗶𝗹𝗱 gave the nicest UI, but also the most complexity - which makes sense with a React front end. 𝗧𝗼𝗱𝗮𝘆’𝘀 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: the shiny interface is great, but the complexity doesn’t disappear - it just moves around (surprise, surprise); also you still need architectural and technical awareness to do anything mildly complex.