Draft Professional Email or Report in Your Voice

Tier: Foundation · Hands-on time: ~25 minutes

Combined with a structured review pass, you get drafts that sound like you and say what you mean.

When to use

• •
  High-stakes or high-effort writing: status reports, client emails, escalations, performance summaries, proposals, incident communications.
• •
  Recurring formats where you write the same shape of document weekly.
• •
  Not for quick internal chat replies — typing those yourself is faster and more authentic.

Prerequisites

• •
  Three samples of your own past writing in the same genre (three past status emails for a status email; don't mix genres). Strip anything confidential from them first.
• •
  The facts for the new piece: bullet points of what actually needs saying.
• •
  Any current AI chat assistant.

The workflow

1. 1.

   Build your voice profile (one-time per genre, reusable forever).

   Copy

   Below are three emails I wrote. Analyze my writing style and produce a
   compact style spec I can reuse: typical greeting and sign-off, sentence
   length, formality level, how I deliver bad news, words and phrases I
   favor, words I never use, how I structure paragraphs, use of bullets.
   Output as a bulleted spec titled "VOICE SPEC", nothing else.
   
   SAMPLE 1: <paste>
   SAMPLE 2: <paste>
   SAMPLE 3: <paste>
   

   Save the VOICE SPEC in your prompt library (pb-05). Edit it by hand where it's wrong — you know your voice better than the analysis does.

2. 2.

   Draft from facts, not from vibes. New chat, then:

   Copy

   Write a <email/report> using this voice spec and ONLY these facts.
   
   VOICE SPEC:
   <paste spec>
   
   AUDIENCE: <who, and what they care about>
   GOAL: <the one thing this must achieve — e.g., get sign-off on the delay>
   FACTS (use all, add nothing):
   - <fact 1>
   - <fact 2>
   - <fact 3>
   
   Constraints: under <N> words. Do not soften or inflate the facts. If a
   fact is unclear, ask me instead of guessing.
   

3. 3.

   Run the review pass. Same chat:

   Copy

   Review your draft against three checks and list issues before fixing:
   1. FACT CHECK — any statement not traceable to my facts list?
   2. READER CHECK — what will this audience misread, object to, or ask?
   3. VOICE CHECK — which sentences would I not actually write?
   Then output the revised draft.
   

4. 4.

   Human pass — the 90-second rule. Read the draft aloud (or subvocalize). Rewrite any sentence you stumble on. Personally verify every number, date, name, and commitment. Then send from your own client, never auto-send.

Tool variants

• Chat assistant (any major vendor): workflow as written. A "project" or "custom instructions" feature lets you store the voice spec permanently so step 1 disappears.
• AI features inside your mail/office suite: convenient for short pieces, but usually can't hold a full voice spec — paste the spec into the instruction box each time, or draft in the chat assistant and paste into mail.
• Local model via Ollama: entirely private path for sensitive drafts (HR, legal-adjacent, client escalations); 8B-class models produce serviceable drafts from a good voice spec, though expect one more revision cycle.

Verification

• •
  Every number, date, name, and commitment in the draft appears in your facts list — check each individually.
• •
  Nothing material was added: the model included no background claims, promises, or apologies you didn't supply.
• •
  The blind-read test: paste your real past email and the AI draft to a colleague — if they can't reliably tell which is which, the voice spec works.
• •
  The GOAL is actually achieved in the text (the ask is explicit, the sign-off requests the decision).

Pitfalls

• Describing style instead of sampling it. "Write professionally but warm" produces generic output; three real samples produce your output. This is the core of the method.
• Fact inflation. Models pad drafts with plausible-sounding commitments ("we'll have this resolved by Friday") you never said. The FACT CHECK pass exists for exactly this.
• Over-hedged bad news. AI drafts soften negative messages into mush. If the point is "the release slips two weeks," verify that sentence survived intact.
• Stale voice spec. Your voice differs by audience (client vs. team vs. exec). Keep one spec per genre, not one global spec.
• Sending without the read-aloud pass. The one AI-sounding sentence you didn't catch is the one people screenshot.

Restricted-environment variant

• •
  Strip client names, deal values, and personal data from both the voice samples and the facts; use placeholders (CLIENT_A, PERSON_1) and substitute back in your mail client before sending.
• •
  Never paste content from a legally privileged thread (legal hold, disciplinary matters) into any hosted tool — draft those with the local-model path or unaided.
• •
  India services context: if your client MSA restricts subprocessors, hosted drafting of client-facing mail may itself be a violation — the local Ollama path (pb-07) makes the whole workflow client-data-safe.

Success criteria

• •
  Drafting time for a recurring format drops from 30+ minutes to under 10, including review.
• •
  Zero factual corrections needed after sending, across five consecutive uses.
• •
  A colleague cannot reliably distinguish your AI-assisted drafts from your unaided ones.
• •
  Your voice spec is versioned in your prompt library and reused across drafts.