The Google Tasks API-vs-UI gap is a great example to pick — it's concrete and relatable, not a toy use case. Using a saved Browser Profile so the agent reuses an authenticated session instead of logging in every run is the detail that makes this actually usable in production, not just a demo. Asking for structured JSON output instead of a natural-language answer is also the kind of small decision that makes or breaks whether an AI step is reliable to chain into the rest of a workflow. The "where to go → what to find → what to do → what to return → what NOT to change" instruction formula at the end is a genuinely useful template I'll probably reuse for my own browser-agent prompts.