You are going to help me build a single reusable file called PROFILE.md. This file will later be loaded into a daily-briefing-and-email-answering assistant connected to my Gmail. Its job is to capture (1) exactly how I write, (2) what matters to me in my inbox vs. what doesn't, and (3) the rules for drafting replies in my voice. Be rigorous - this file's quality determines whether the assistant sounds like me or like a generic bot.
Work in three phases. Do not skip ahead, and do not write PROFILE.md until Phase 3.
Phase 1 - Mine my actual voice (use the Gmail connector)
Pull a representative sample of emails I have sent (aim for 30-50 across the last few months, spanning different recipients: investors, teammates, customers/partners, vendors, personal). Then analyze and report back to me on:
Sentence rhythm & length - do I write short and punchy, long and detailed, or mixed by context?
Punctuation fingerprints - em-dashes, ellipses, line breaks, ALL-CAPS for emphasis, exclamation use, emoji use.
Openers & closers - how I actually greet and sign off (by relationship type).
Tone register - warm/casual vs. formal, and how it shifts between, say, an investor and a teammate.
Recurring phrases & filler - words or constructions I lean on.
How I make asks - direct CTA vs. soft CTA, and how I soften or sharpen requests.
What I never do - patterns conspicuously absent from my writing.
Show me a concise summary of these findings and flag anything that surprised you. Quote a few short representative snippets so I can confirm they sound like me. Wait for my confirmation or corrections before continuing.
Phase 2 - Interview me on priorities and rules
Ask me these, one cluster at a time (don't dump all at once). Use what you already learned from my inbox to make smart guesses, then have me confirm or adjust rather than answering from scratch:
VIP senders / domains - whose email always rises to the top? (people, companies, domains)
High-signal topics - subjects/keywords that mean "urgent" or "important" regardless of sender (e.g. fundraising, a specific deal, anything from my kids' school).
Noise - what's safe to summarize-and-skip or batch (newsletters, receipts, cold outreach, notifications).
Hard "do not auto-handle" list - categories where the assistant must never draft or send without me (legal, anything money-moving, sensitive personal, board/investor commitments).
Response-time expectations - what needs same-day vs. what can wait.
Reply archetypes - the recurring email types I answer and how I want each handled (e.g. "intro request -> warm, brief, offer two times"; "vendor pitch -> polite decline"; "teammate question -> direct answer, no fluff").
Draft vs. send - should the assistant only ever draft, or auto-send certain low-risk categories?
Daily briefer shape - when I want it, how long, and what sections (e.g. "Needs me today," "FYI," "Drafted & waiting," "Ignored - here's why").
Phase 3 - Write PROFILE.md
Synthesize Phases 1-2 into one clean, well-structured markdown file with these sections:
About Me - one tight paragraph of identity/context so the assistant answers intelligently (who I am, my company, my role, what I care about).
Voice & Style Rules - concrete, testable rules ("Use ellipses for trailing thoughts... not commas." / "Sign off to teammates with just my first name."). Include a short DO / DON'T list.
Tone Matrix - a small table mapping recipient type -> tone, opener, closer, length.
Email Triage Taxonomy - Priority tiers (P1/P2/P3/Noise) with the rules and example senders/topics that sort into each.
Reply Playbook - the archetypes from Phase 2, each with a one-line strategy + a short sample reply written in my voice.
Guardrails - the never-auto-send list and the draft-only vs. auto-send policy, stated unambiguously.
Daily Briefer Spec - exact format, sections, length, and tone for the morning brief.
Calibration Examples - 2-3 full sample emails (a good one and a deliberately wrong one) so a future assistant can tell "sounds like me" from "doesn't."
Write it so a fresh Claude instance with zero context could read only this file and convincingly operate my inbox. Output it as a single copy-pasteable markdown block.
Show full prompt
You are going to help me build a single reusable file called PROFILE.md. This file will later be loaded into a daily-briefing-and-email-answering assistant connected to my Gmail. Its job is to capture (1) exactly how I write, (2) what matters to me in my inbox vs. what doesn't, and (3) the rules for drafting replies in my voice. Be rigorous - this file's quality determines whether the assistant sounds like me or like a generic bot.
Work in three phases. Do not skip ahead, and do not write PROFILE.md until Phase 3.
Phase 1 - Mine my actual voice (use the Gmail connector)
Pull a representative sample of emails I have sent (aim for 30-50 across the last few months, spanning different recipients: investors, teammates, customers/partners, vendors, personal). Then analyze and report back to me on:
Sentence rhythm & length - do I write short and punchy, long and detailed, or mixed by context?
Punctuation fingerprints - em-dashes, ellipses, line breaks, ALL-CAPS for emphasis, exclamation use, emoji use.
Openers & closers - how I actually greet and sign off (by relationship type).
Tone register - warm/casual vs. formal, and how it shifts between, say, an investor and a teammate.
Recurring phrases & filler - words or constructions I lean on.
How I make asks - direct CTA vs. soft CTA, and how I soften or sharpen requests.
What I never do - patterns conspicuously absent from my writing.
Show me a concise summary of these findings and flag anything that surprised you. Quote a few short representative snippets so I can confirm they sound like me. Wait for my confirmation or corrections before continuing.
Phase 2 - Interview me on priorities and rules
Ask me these, one cluster at a time (don't dump all at once). Use what you already learned from my inbox to make smart guesses, then have me confirm or adjust rather than answering from scratch:
VIP senders / domains - whose email always rises to the top? (people, companies, domains)
High-signal topics - subjects/keywords that mean "urgent" or "important" regardless of sender (e.g. fundraising, a specific deal, anything from my kids' school).
Noise - what's safe to summarize-and-skip or batch (newsletters, receipts, cold outreach, notifications).
Hard "do not auto-handle" list - categories where the assistant must never draft or send without me (legal, anything money-moving, sensitive personal, board/investor commitments).
Response-time expectations - what needs same-day vs. what can wait.
Reply archetypes - the recurring email types I answer and how I want each handled (e.g. "intro request -> warm, brief, offer two times"; "vendor pitch -> polite decline"; "teammate question -> direct answer, no fluff").
Draft vs. send - should the assistant only ever draft, or auto-send certain low-risk categories?
Daily briefer shape - when I want it, how long, and what sections (e.g. "Needs me today," "FYI," "Drafted & waiting," "Ignored - here's why").
Phase 3 - Write PROFILE.md
Synthesize Phases 1-2 into one clean, well-structured markdown file with these sections:
About Me - one tight paragraph of identity/context so the assistant answers intelligently (who I am, my company, my role, what I care about).
Voice & Style Rules - concrete, testable rules ("Use ellipses for trailing thoughts... not commas." / "Sign off to teammates with just my first name."). Include a short DO / DON'T list.
Tone Matrix - a small table mapping recipient type -> tone, opener, closer, length.
Email Triage Taxonomy - Priority tiers (P1/P2/P3/Noise) with the rules and example senders/topics that sort into each.
Reply Playbook - the archetypes from Phase 2, each with a one-line strategy + a short sample reply written in my voice.
Guardrails - the never-auto-send list and the draft-only vs. auto-send policy, stated unambiguously.
Daily Briefer Spec - exact format, sections, length, and tone for the morning brief.
Calibration Examples - 2-3 full sample emails (a good one and a deliberately wrong one) so a future assistant can tell "sounds like me" from "doesn't."
Write it so a fresh Claude instance with zero context could read only this file and convincingly operate my inbox. Output it as a single copy-pasteable markdown block.
