# Interview Guide

The interview mixes context mining, open-ended questions, and interactive choice questions. Small batches, 3 to 5 per turn. Where the environment provides an interactive question interface (such as the AskUserQuestion tool), use it for every single-choice and multiple-choice question; fall back to asking in chat where it does not exist. Open-ended questions always go in chat.

## Step 0: Mine before you ask (mandatory, before ANY interview question)

For every interview area below, first check what you already have. Two sweeps, done at the start and reported to the user:

1. **Memory sweep.** Evaluate everything memory holds about how this user has used Claude: stored facts, preferences, past corrections, recurring tasks, prior conversations where available. Summarize what you believe you already know about them and how they like to work. Present it for confirmation before relying on it; memory can be stale.
2. **Connector sweep.** Enumerate the connectors and tools actually available in this session (email, calendar, document drives, chat tools, CRM, and anything else). Tell the user what is connected and what you intend to pull from each, then pull with their go-ahead:
   - **Email (highest value):** a handful of recent SENT emails. Sent mail is the best voice sample that exists: it is how they write when nobody is watching. Read for rhythm, greetings, sign-offs, formality, humor, directness; report observations, not contents.
   - **Documents:** anything they authored (memos, decks, posts) for long-form register.
   - **Chat tools:** their casual register, which often differs sharply from email.
   - **Calendar:** role rhythm, the audiences they meet, meeting load (informs what Claude will be drafting).
   - **CRM or company systems:** role, title, company context.
3. Where a connector is not linked or a sweep comes up empty, say so plainly and ask the user to paste 3 to 5 recent sent emails or writing samples instead. Never skip voice samples; the persona is built on them.

## Step 1: Company grounding

If the sweeps did not establish the company, ask for the company website first, before any company questions. Then research deeply: the site, web search, the 10-K if public, news, and any connected sources. Present the company picture (what it does, customers, positioning, the words the company itself uses) for confirmation or revision, and state what could not be found online. Only then ask the residual company questions research could not answer.

## Step 2: Identity and role (open + choice)

- Open: describe your role in your own words. What does a great week look like? What do you spend too much time on?
- Choice (interactive): what will Claude mostly do in your name? (Emails and replies / Documents and memos / Decks and talks / Analysis and research / Code / Social posts; multi-select.)
- Choice: who are your main audiences? (Board and investors / Executives and peers / Your team / Customers and partners / Public; multi-select.)
- Open: names, titles, and relationships Claude should know (assistant, co-founder, board chair). Pronouns and anything commonly mispronounced or misspelled.

## Step 3: Voice observation (from the samples)

Analyze the gathered samples and present an observed voice profile before asking anything about voice: typical sentence length and rhythm, formality by audience, greetings and sign-offs actually used, punctuation habits, emoji or none, words and constructions they favor, how they deliver a no, how they praise. Ask the user to confirm, correct, or contest. This ordering matters: people describe their voice inaccurately; the samples do not.

## Step 4: Voice probes (the calibration drills)

Run 3 to 5 short drafting probes across DIFFERENT registers. For each: produce two or three candidate replies at different calibrations (for example, warmer vs. flatter; brisker vs. fuller), then ask the user to pick the closest AND edit a line or two. Their edits are the strongest signal in the whole interview; mine every edit for a rule (what they cut, what they added, what they refused).

**Print the drafts before the choice, always.** The full text of every candidate must appear in the chat message BEFORE the interactive choice interface is shown, clearly labeled (Candidate A, Candidate B) with a one-word calibration tag. The choice interface then carries only the short labels, and, where the interface supports option descriptions, a one-line excerpt of each draft. Never present a choice whose substance exists only inside the choice options: choice interfaces truncate, and a user shown "Candidate A (fuller)" with no draft in view cannot choose. If for any reason the drafts were not printed first, print them and re-ask.

The probe set, one per register:
1. **Decline gracefully:** an invitation or request they must turn down without burning the bridge.
2. **Bad news to a customer or partner:** something went wrong; own it in their voice.
3. **Praise to a teammate:** publicly or privately; watch how effusive they allow themselves to be.
4. **The terse acknowledgment:** the two-line reply to something routine; watch the floor of their formality.
5. **The stress test (extreme register):** a theatrical, over-the-top inbound (an unhinged multi-page complaint from a self-important counterpart, with absurd demands and legal bluster) that they must answer. This one earns its place because it reveals what the mundane probes cannot: how much steel they show under attack, whether humor is permitted when the stakes are real, how they handle someone else's drama without absorbing it, and where the line is between firm and combative. Invent a vivid scenario appropriate to their industry; make it fun, but make the reply matter.

## A note on the single-prompt approach (and why this skill goes further)

A popular technique is to hand Claude one colorful inbound email and say: draft the reply as me. It works better than nothing, and the extreme-register version of it is genuinely useful, which is why probe 5 exists. But used alone it has four failures the probe set is designed to fix: it captures one register and most writing lives in the mundane ones; it overfits to drama, producing personas that are all steel and wit; it tests output only, learning nothing about preferences, boundaries, or decision defaults; and it offers no calibration baseline, so the user cannot say warmer or flatter, only yes or no. Multiple probes across registers, with candidate drafts and user edits, produce a persona the single prompt cannot.

## Step 5: Preferences, defaults, and boundaries (open + choice)

- Formatting: bullets or prose, email length ceiling, headers or none, bold usage. Choice questions work well here.
- Language rules: words and phrases to always use and to never use (many executives ban specific corporate cliches; capture them verbatim). Punctuation rules (em dashes, exclamation marks, emoji policy).
- Decision defaults: when drafting on their behalf, should Claude default to committing (yes with conditions) or hedging (options presented)? What may Claude decide autonomously and what must always come back as a question?
- Sign-off inventory: which closings for which audiences.
- The red lines: topics, commitments, or tones Claude must never produce in their name.

## Step 6: Confirm the assembled persona

Assemble the persona.md from the template, present it section by section with a one-line explanation of what each section does, and loop on revisions. Then distill the custom-instructions excerpt and confirm it too. Their approval gates delivery.