Every underwriter we've worked with has a tell. One always opens external emails with "Hi Joe — quick note —". Another never uses contractions when corresponding with brokers. A third has a specific way of declining a binder bump that's polite without leaving any wiggle room.

An AI that can't replicate those quirks produces drafts that read as foreign. They get rewritten anyway. Net productivity gain: zero.

How email-style-learn works

The skill samples your last N sent emails from Outlook (default 50) and extracts:

• Salutation patterns by recipient type (external broker vs internal teammate vs senior leadership)
• Sign-offs, with frequency
• Average sentence length and paragraph structure
• Vocabulary quirks — frequent phrases, hedges, industry jargon you actually use
• Whether you bullet-list action items at the end or weave them into prose
• Standard footer / signature block
• Words you demonstrably avoid

The output is saved as ~/.insightworker/playbooks/email-style.md — a markdown file you can read, edit, and version-control. No black box.

How email-draft-reply uses it

When you ask for a draft reply, the agent loads your style guide, the source email, and any relevant industry playbook (e.g. an insurance glossary). It drafts to a markdown file under drafts/email/ — never auto-sent. You review, tweak, and send through Outlook (or have the agent send via send_email, which always permission-prompts).

If the source email is flagged as legal, regulatory, or contains complaint / subpoena / litigation language, the skill refuses to draft a substantive reply and routes it to compliance review with a one-line acknowledgement instead.

Why this matters

Underwriters and account managers spend a lot of time replying to emails that need their judgement but not their time. A draft that sounds like them is a draft they'll actually send. That's where the productivity win lives — not in generating more drafts, but in generating drafts good enough to ship after a 30-second read.