Interview questions

Open this at step 6, after the counting. Ask three or four at a time. After each group, play back what you heard in a few lines and let the user correct it. People correct a paraphrase far more readily than they volunteer detail.

Skip any question the sent mail already answered. Ask it as a confirmation instead: "Your mail shows X. Is that right?"

Who and how

1. Who do you write to most in a normal week? Name the kinds of people, not individuals.
2. Which of those need a different tone from you? What changes?
3. Is there anyone you write to more formally than your mail shows, or less?
4. What does a good reply from you do for the other person? Listening for: speed, clarity, options, warmth.

What the mail cannot show

5. Which links do you send people to book time, and which one goes to whom? Paste them. These become the only links the skill may use.
6. Do you put prices in email, or only on calls? How do you round them?
7. Are there words or phrases you would never send? Anything an AI draft wrote that made you wince?
8. When someone cancels, apologises or goes quiet, what do you do?
9. When you say no to something, how do you say it?
10. What do you want people to be able to do with your email afterwards? Listening for: forward it internally, act without a call, reply in one line.

About the drafts

11. Think of the last AI draft you rewrote. What was wrong with it?
12. What would make you send a draft without touching it?
13. Is there anything rough about how you write that you want kept? Typos, long sentences, asides.

Close

14. What else should a person writing as you know that we have not covered?

Repeat the close until the answer is nothing.

Where the answers go

Answers	Land in
1 to 4	the register map
5	the verified-links list in the draft mechanics
6, 8, 9	patterns, each with its register
7, 11	the avoid list
10, 12	what good looks like, at the top of the skill
13	the "keep the rough edges" rule

Measurement checklist

Open this at steps 2 and 4. It covers which emails go in the sample and what to count across them. Every line here ends as a number in the finished skill.

Building the sample

Take 30 to 40 sent emails from the last one to two months. Older mail drifts: roles, clients and habits change.

Leave out:

• Auto-replies, calendar notices, receipts and anything sent by a tool on the user's behalf.
• Forwards with no added text, and replies that are one word ("Thanks!").
• Emails that are mostly quoted text from someone else.
• Mail written by an assistant or a shared inbox.

Keep the one-line replies that say something ("No problem at all, talk Thursday"). They are the most common email most people send, and leaving them out makes the voice look longer and more formal than it is.

Spread the sample across recipients. Aim for at least three emails in each kind the user sends often:

Kind of recipient	Examples
Client, active work	updates, plans, deliverables, scope changes
Prospect	follow-up after a call, proposals, pricing
Someone they coach or manage	check-ins, feedback, missed meetings
Partner or boss	contract changes, reporting, asks
Introductions	ones they make, ones they receive, replies to the person who introduced them
Vendor or admin	support requests, billing, platform issues
Short replies inside a live thread	scheduling, reactions, quick answers

Under 20 usable emails: mark every rule provisional and re-sweep when there are 30.

What to count

Measure	How to record it	What it usually decides
Length	words per email: median, and the share under 50 and over 300	the default length per register
Greeting	each form with its count, including no greeting at all	the greeting rule, per register
Closer line	each line before the sign-off with its count	a short list in order of frequency
Sign-off	first name, full name, initials, none, with counts	when the full name appears (often vendor mail only)
Exclamation points	per email: zero, one, two to three, more	whether warmth is the default or the exception
Dashes	long dashes per email; short dashes in ranges	whether a dash in a draft is a tell
Paragraph length	sentences per paragraph, words per sentence	how dense a draft may get
Lists	numbered versus bulleted, and what each is used for	often: numbers when each item needs an answer
Bold	where it appears	labels, the one real ask, or never
Repeated phrases	any phrase in three or more emails, with its count	the phrases the skill should reach for
Advice verbs	"I'd", "you should", "I recommend", with counts	how the person gives advice
Asks	how a request is worded and whether it carries a softener	the ask pattern
Apologies	the first line after someone else apologises or cancels	a reflex the skill should copy exactly
Money	whether prices appear in email, how they are rounded	the price rule
Typos and run-ons	count them, do not fix them	a reason not to polish
Links	which booking or document links appear, and where	the verified-links list, confirmed in the interview

Writing findings

Write each finding with its count and its scope: "Greeting is 'Hi' or 'Hey' plus a first name in 38 of 40. 'Hey' goes to people they know, 'Hi' to new contacts and groups." Then quote one line from the mail as the example. A finding with no count is an impression, and impressions are what the skill exists to replace.

Rewrite diff worksheet

Open this at step 3, and again every time the user edits a draft after the skill is live. A rewrite shows what the person refuses to send, which the sent mail alone cannot show.

Finding the rewrites

• Ask the user for two or three AI drafts they edited, with the version they sent.
• Or search their drafts and sent mail for near-identical pairs written within a day of each other.
• Or skip this step and pick it up at step 9, where the test drafts become the first rewrites.

Diffing one rewrite

Go line by line and sort every change into one of five columns:

Change	What to record	Example of what it becomes
Cut	the exact words removed	an avoid-list entry
Added	the exact words added	a pattern, or a phrase to reach for
Softened	before and after	an ask or a deadline rule
Hardened	before and after	where the person is blunter than the AI
Moved	what moved and where	an order rule (commitments before asks, price early)

Then answer three questions about the whole email:

1. What kind of recipient was it? The rule belongs to that register until the sent mail shows it holds elsewhere.
2. Did the person give a reason, in the edit or out loud? A stated reason outranks an inferred one.
3. Did they leave something rough on purpose: a typo, a run-on, a repeated phrase? Record it as kept, so the next draft does not smooth it away.

Turning the diff into the skill

• A change seen once is a pattern for that register, marked with how many times it has been seen.
• A change seen in two different registers becomes a general rule.
• When a rewrite and the sent mail disagree, the sent mail wins on frequency. The rewrite still wins for the exact kind of email it came from.
• Every change goes into the patterns section. A learning-log line alone is not read at draft time, so it does not change the next draft.

What diffs usually turn up

These come up often enough to check for on purpose:

• The AI draft was longer than anything the person sends to that recipient.
• The AI used a word the person never uses ("leverage", "seamless", "unlock").
• The AI wrote an ask as an instruction where the person writes it as an offer.
• The AI committed someone else's time ("my colleague and I will review it").
• The AI apologised at length where the person apologises in three words.
• The AI used a link from the signature or an old thread that was not the right one.
• The AI polished away a habit the person keeps on purpose.

Voice skill template

Open this at step 8. It is the shape of the skill this build produces. Fill every section from the counts, the diffs and the interview. Save it in the user's own skills folder under a name like <their-name>-email.

Frontmatter

---
name: <their-name>-email
description: Load whenever drafting an email for <name>: client updates, follow-ups, introductions, coaching, partner and vendor mail, or any message sent from their inbox. Not for posts or newsletters.
---

Sections, in this order

1. Evidence base

One paragraph: how many sent emails and how many rewrites the skill is built on, the date range, and which source wins when they disagree.

2. Draft mechanics

Copy these in as written. They are the rules that stop a voice skill from doing damage:

• Drafts only, always. Never send, even when asked to "send them". The person sends everything themselves.
• A reply stays in its original thread, created as a reply to the message it answers.
• Editing a reply draft can detach it from its thread in some mail clients. After any edit, check the draft still sits in the thread. If it detached, delete it and recreate it as a reply.
• Never insert a link that is not on the verified-links list. Write insert link here instead.
• A link in the email signature is not a booking link unless the person says it is.
• Never say a file is attached unless it is. Write [attach file here]. A file the agent built and attached in the same step counts as verified.
• The verified-links list: each link, what it is for, and who it goes to.

3. Register map

A table with one row per register: who, median length, warmth, greeting, closer, exclamation range. Then the median length across all mail, and the share under 50 words. Pick the register before writing a word.

4. Mechanics of the page

Greeting, paragraph length, lists, bold, sign-off, closer lines in order of frequency, dashes, exclamation points. Each with its count.

5. Patterns

Numbered. Each pattern has:

• a name the person would recognise,
• the register it applies to,
• the count behind it (sent mail) or the rewrite it came from,
• one quoted line from their real mail.

Common pattern areas to check for: how they open after someone apologises, how they ask for things, how they give advice, how they say no, how they show proof, how they handle money, how they make and answer introductions, how they schedule.

6. Avoid list

Words, phrases and moves the person cuts, each with where it was seen. Include "over-polishing" if their real mail has typos and run-ons.

7. Examples

Three to five real emails, one per register, with names and private details changed if the skill will be shared. Under each: why it is a good example.

8. How to grow this skill

• Every rewrite: diff it, add the change to Patterns with its register, then log it.
• Every quarter: re-sweep 30 to 40 sent emails and correct any rule the counts no longer support.
• A rule from one email type stays scoped to that type until the sent mail shows it holds elsewhere.

9. Learning log

A table: date, source (rewrite or sweep), what changed in the skill.