Build Your Personal Assistant Operating System

3.3

Teach the Assistant Your Writing Voice

AI-generated text sounds like everyone and no one

The assistant can write fluently. That is the easy part. The hard part is getting it to write like you. Without voice guidance, every draft comes out in the same polished, generic register: grammatically correct, tonally flat, and indistinguishable from anything else an AI produced.

Your writing voice is a pattern of choices you make consistently. Sentence length, vocabulary range, level of directness, how you open an email, how you handle disagreement, which words you never use, and which phrases you always reach for. These choices are learnable, and the assistant can learn them if you give it the right material.

This chapter teaches you to build a voice profile: a document that captures your writing patterns in enough detail that the assistant's drafts sound like you wrote them on a busy day. The profile works across every that produces text, from email replies to reports to meeting summaries.

A voice profile teaches the assistant the specific patterns that make your writing yours.

Writing Voice Profile

Promise	A reusable style document that teaches the assistant your writing patterns: sentence structure, vocabulary, directness, audience adaptation, and banned patterns.
Sources	Ten or more examples of text you have written across different audiences. Corrections you have made to assistant drafts. Your explicit preferences about tone, formality, and word choice.
Output	A voice profile with five dimensions described, correction pairs showing wrong and right examples, banned patterns, and a style test the assistant can run against its own drafts.
Saves to memory	The voice profile, saved as a shared resource. Every that produces text references this profile.
Approval boundary	The assistant uses the profile to draft. You review and approve. The profile never overrides your edits; your edits update the profile.
Failure behavior	If the profile is incomplete, the assistant notes which dimensions are missing and drafts conservatively, matching the examples it has rather than guessing about gaps.
Review criteria	A draft written with the profile should be close enough to your voice that a colleague could not tell whether you or the assistant wrote it.
Graduation trigger	When you stop rewriting drafts for tone and only edit for content, the profile is working.
Privacy tier	Standard. The profile itself contains your writing patterns and example text.

Five dimensions capture what makes your writing yours

Voice is not one thing. It breaks down into dimensions that the assistant can learn independently. Each dimension is a spectrum, and your position on that spectrum is consistent across your writing.

Five dimensions of writing voice

Dimension	What it captures	Example position
Sentence architecture	Average length, variation, how you combine clauses	Short sentences for emphasis. Longer sentences when explaining . Rarely more than two clauses.
Vocabulary register	Formal vs. casual, technical vs. plain, how you handle jargon	Casual with clients, direct with the team, precise with technical terms but always defined on first use.
Directness	How quickly you get to the point, how much you provide before the ask	Lead with the request. Provide after. Skip preamble.
Warmth	How personal, empathetic, or transactional your tone is	Warm with individuals. Matter-of-fact in group communications. Always acknowledge the other person's situation before making a request.
Signature patterns	Phrases you use often, ways you open and close, distinctive habits	Always open client emails with their name, no 'Hi.' Close with 'Let me know if questions come up.' Never use 'please advise.'

The assistant does not need to replicate every quirk. It needs to match the dimensions closely enough that the result feels like your voice. The profile is a compass, not a transcript.

Correction pairs teach voice faster than descriptions

Describing your voice is difficult. Showing it is faster. A correction pair is a before-and-after example: the assistant wrote this, but you would write it this way instead.

Three correction pairs that teach more than a paragraph of description

Assistant wrote: 'I wanted to reach out regarding the upcoming project timeline and share some thoughts on how we might approach the deliverables.'
You would write: 'Quick update on the Atlas timeline. Two things changed this week.'
What the pair teaches: lead with the subject, skip the preamble, name the project.
---
Assistant wrote: 'Thank you for your patience while we worked through the details of this engagement.'
You would write: 'Thanks for sticking with us on the details. Here is where we landed.'
What the pair teaches: informal gratitude, then move forward. No corporate phrasing.
---
Assistant wrote: 'Please do not hesitate to contact me if you have any further questions or concerns.'
You would write: 'Let me know if questions come up.'
What the pair teaches: shorter closings, conversational tone, your sign-off phrase.

Ten correction pairs teach the assistant more about your voice than a page of description. Each pair is a concrete rule: when you see this pattern, replace it with this one. The assistant applies these rules to every draft.

Banned patterns prevent the most common voice failures

Some AI writing patterns appear so consistently that banning them is more effective than correcting them one at a time. Common bans include:

Corporate filler: 'I wanted to reach out,' 'just circling back,' 'per our conversation,' 'please do not hesitate.'
Hedge stacking: 'I think it might be possible that perhaps we could consider...' when you would just say 'We should try X.'
False enthusiasm: 'Excited to share,' 'thrilled to announce,' 'really looking forward to' when your actual tone is calm and direct.
Passive deflection: 'It was decided that' when you would say 'I decided' or 'the team decided.'

Your banned list will be personal. Some writers ban exclamation points. Some ban 'utilize' in favor of 'use.' Some ban rhetorical questions in professional communication. Each ban removes a pattern that makes the assistant's drafts sound generic.

The style test catches voice drift before you send anything

After building the profile, test it. Give the assistant a writing task across three audiences: a client email, a team message, and a report paragraph. Review the output against your voice dimensions. If the client email is too formal, the team message too cautious, or the report too casual, adjust the relevant dimension.

Run this test monthly. Voice profiles drift as the assistant accumulates instructions from other conversations. A monthly style test, using the same three prompts, catches drift before it compounds.

How this module breaks

What goes wrong	How you notice	What to correct once	What rule to save
The assistant matches your casual voice perfectly but uses the same register in a formal report.	A client report reads like an internal Slack message because the voice profile emphasized your casual style without audience separation.	Add audience layers to the profile. Each audience gets a modifier for formality, warmth, and directness. The base voice stays consistent; the register shifts by .	The voice profile includes at least two audience layers (professional and internal). Each layer adjusts formality, warmth, and directness relative to the base voice.
The profile captures your current voice but not how it changes over time.	Six months later, the assistant writes like you did last year. Your style has evolved but the profile has not.	Schedule a quarterly voice review: re-run the style test with fresh examples. Update correction pairs with recent drafts. Retire bans that no longer apply.	The voice profile is reviewed quarterly. New correction pairs from recent writing replace stale examples. Banned patterns are confirmed or removed.
The banned-pattern list is so long that the assistant writes in a constrained, unnatural way.	Drafts are technically correct but feel stiff because the assistant is avoiding so many patterns that it has no room to write naturally.	Limit bans to ten or fewer patterns. Prioritize the bans that cause the most noticeable voice failures. General guidance works better than exhaustive prohibition.	The banned list stays under ten patterns. Each ban must address a recurring failure that corrections alone cannot fix.

Your turn

Build your writing voice profile

You'll produce a voice profile with five dimensions, ten correction pairs, a banned-pattern list, and a style test you can run monthly.

Why this exercise matters

Without a voice profile, every that produces text sounds generic. The profile is what makes the system sound like you.

You’ll leave with

A voice profile covering sentence architecture, vocabulary register, directness, warmth, and signature patterns.
Ten correction pairs showing the assistant's default vs. your preferred phrasing.
A banned-pattern list of no more than ten items.
Audience layers for at least two contexts (professional and internal).
A style test with three prompts you can run monthly.

Use the prompt in order

1
Share your writing examples and let the assistant analyze your patterns
Paste the ten examples and tell the assistant which audience each was written for. Ask it to identify your patterns across the five dimensions: sentence architecture, vocabulary register, directness, warmth, and signature patterns.
2
Correct the analysis where it misses your intent
The assistant will describe your voice. Some descriptions will be accurate. Some will miss the reason behind a pattern. For example, if you write short emails to clients because you respect their time, correct the interpretation so the assistant knows the motive behind the brevity.
3
Build ten correction pairs from recent assistant drafts
Find ten places where the assistant wrote something you rewrote. For each, record the original and your version. These pairs are the fastest way to teach specific voice adjustments.
4
Run the style test and refine
Ask the assistant to write three things using the profile: a client email, a team update, and a report paragraph. Compare each to what you would have written. Adjust the dimensions and correction pairs based on what was off.

Prompt details

Each detail will be inferred, or you will be asked to clarify in the chat.

1/1 details

1 detail will be inferred, or you will be asked to clarify in the chat

Starter prompt text

Open the full text if you want to check what will be copied.

Show starter prompt

Analyze my writing voice from these examples and build a voice profile I can reuse across all my assistant modules.

My writing examples (labeled by audience):
Infer from what you know about me or ask if it is unclear

For each example, identify patterns in:
1. Sentence architecture (length, variation, clause structure)
2. Vocabulary register (formal vs. casual, jargon handling)
3. Directness (how fast I get to the point)
4. Warmth (personal vs. transactional tone)
5. Signature patterns (phrases I use often, openings, closings)

After the analysis, build:
- A five-dimension voice profile
- Audience layers for professional and internal communication
- A list of patterns I never use (banned list)

Then test the profile: write a client email, a team update, and a report paragraph using my voice. I will compare and correct.

What the answer should give you

A voice profile with: sentence architecture (short and direct, rarely more than two clauses), vocabulary register (casual with clients, shorthand with team, precise in reports), directness (lead with the request, context after), warmth (acknowledge the person before the ask), signature patterns (opens with first name, closes with 'let me know if questions come up'). Banned list: 'please advise,' 'just circling back,' 'excited to share.' Audience layers for client and team. Three test outputs for comparison.