If we want to build a smaller version of an internal AI/automation system, where should we start?

Start with basic infrastructure: Vercel, GitHub, and Supabase, which the speaker found user-friendly and sufficient to get going. Early on, the process was as simple as screenshotting setup steps (e.g., 'sign up for GitHub') and asking Claude what to do next, screenshot by screenshot. Build a simple site/tool structure, then pick one small, manual pain point to automate first — for example, a bank statement renamer that takes an uploaded PDF and renames it, without even connecting to SharePoint at first. The approach is to describe the problem to Claude, provide the inputs (e.g., a CSV from a client and a CSV export from QuickBooks), describe the desired end result (e.g., posting a journal entry or bill), and let it figure out how to make the two talk to each other.