Onboarding personal-context tools: derive everything, ask only for guardrails

When designing onboarding for a tool that needs personal/work context, a setup form is the failure mode — users abandon at the blank page. Almost everything is derivable from 90 days of mail and calendar: domains from which account received what, org membership from counterparty email domains, who matters from contact frequency and reply latency, and notably per-domain writing tone from the users own sent folder. The only thing with no signal is guardrails (what the agent must never do unattended), so that is the sole thing worth a form and it is four questions. For ongoing nuance, reuse the approval gate as the learning channel: when a user edits a draft before approving, that diff is the training signal — surface the inferred rule in plain English with a one-key reject so bad inferences die before hardening. Design principle: the user should only ever correct, never author.