For Riccardo

Articles that came out of our calls — things I said I’d look into or wasn’t sure about, checked against primary sources. Each ends with what was verified, when, and what stays inferred.

• What actually protects you when an agent has your files — and how real is prompt injection?
• When does routing a bounded task to a cheaper model via API pay off?
• What do Claude’s permission modes actually gate, from always-ask to skip-everything?
• What do Claude’s plan tiers buy, and when does overage beat upgrading?
• How do you get AI output into Google Docs and Sheets — formatting and version history intact?
• Can a scheduled local Claude Code run use your connectors, or does it need the cloud?