How to Build a "Communication Profile" That Makes AI Write Exactly Like You

Sentence Cadence and Structure

This is the skeleton. Track your average sentence length, the ratio of short declarative sentences to longer compound ones, and whether you use fragments intentionally. Some writers naturally alternate between punchy two-word sentences and elaborate 30-word constructions. Others stay in a narrow band. The model needs to know which you are.

Greetings and Sign-offs

This sounds trivial. It isn't. Your opener and closer are the two elements that people read most carefully. Do you write "Hi Sarah," or "Sarah -"? Do you close with "Best," or "Talk soon," or nothing at all? These patterns are fingerprints.

Vocabulary Preferences

Every writer has signature words. You might overuse "essentially" or lean on "the issue is" as a transition. You might avoid jargon entirely or use it as a trust signal with specific audiences. A good Communication Profile catalogs these patterns explicitly - including abbreviations, contractions, and any informal constructions you favor.

Grammar and Formatting Habits

Do you use em dashes or parentheses? Oxford commas or not? Do you write in short paragraphs (2–3 sentences) or longer blocks? How often do you use bullet points? Exclamation marks? The model will replicate whatever pattern you show it - but only if you've documented the pattern clearly enough.

Formality Spectrum

Place yourself on a scale. Not "formal" or "casual" - those are too broad. Something like: "Professional-warm. Uses first names immediately. Avoids corporate phrasing but maintains clear authority. Occasionally self-deprecating. Never uses sarcasm in client-facing communication."

Persuasion and Rhetoric Style

This is the dimension that separates a Communication Profile from a generic style guide. How do you actually move people to action? Do you build evidence first and conclude with a recommendation? Do you lead with the ask and justify afterward? Do you frame things as questions to create buy-in, or state positions directly? Your persuasion style is as distinctive as your handwriting, and it's the hardest dimension for an AI to infer without explicit documentation.

The Extraction Process: How to Build Your Profile

You need 10–15 samples of your own writing. Emails work best because they're natural, unedited, and represent your actual voice rather than your "published" voice. Select samples that cover different contexts - a quick internal update, a client-facing proposal, a difficult feedback conversation, a casual follow-up.

Then run them through this extraction prompt:

Analyze the raw writing samples below across these dimensions:

1. Sentence Cadence & Structure: Track average sentence length, variety in length, and the ratio of simple to compound/complex sentences.
2. Greetings & Sign-offs: Identify the exact vocabulary, level of intimacy, and formatting used for starting and ending messages.
3. Vocabulary Preferences: Note signature words, repetitive verbs/adjectives, jargon vs. simple terms, and any abbreviations.
4. Grammar & Formatting: Check capitalization habits, punctuation patterns, paragraph lengths, and bullet usage.
5. Formality & Distance: Place the author's voice on a spectrum from highly formal/transactional to warm/informal/intimate.
6. Persuasion & Rhetoric: Identify how the author frames requests, handles objections, or guides the reader to action.

Output a structured document labeled "COMMUNICATION PROFILE" containing your findings. The profile must be detailed enough that another AI model could accurately reproduce the writing style using only this document.

=== WRITING SAMPLES ===
{{email_samples}}

That last line - "detailed enough that another AI model could accurately reproduce the writing style" - is doing critical work. It forces the model to be specific rather than impressionistic. Without it, you'll get vague summaries like "professional and friendly." With it, you'll get operationally useful parameters.

Author's Note: I've tested this extraction across ChatGPT, Claude, and Gemini. Claude tends to produce the most granular profiles - likely because of its strength with long-document analysis. But all three produce usable output. The quality depends more on the diversity and quantity of your writing samples than on the model you use for extraction.

The Persistence Problem (And How to Solve It)

The most common failure point in voice cloning isn't profile quality - it's context expiration. LLMs operate under a stateless paradigm: every new chat session flushes the context window completely. ChatGPT, Claude, and Gemini will not carry your Communication Profile across separate conversations unless you architect persistence into your workflow.

Three solutions, in order of increasing robustness:

• Copy-paste method. Save your Communication Profile as a markdown file (My_Email_Style_Guide.md). At the start of any new conversation, paste it in with the instruction: "Use this Communication Profile for all writing in this conversation." Simple, portable, works everywhere.

• Platform-native persistence. ChatGPT's Projects feature lets you attach files to a project's knowledge base. Claude's Projects work similarly. Gemini offers Gems with saved system instructions. In each case, you upload the profile once and it persists across conversations within that project. This is the most frictionless option for daily use.

• System prompt integration. If you're working through an API or building automated workflows, embed the Communication Profile directly in the system prompt. This is the most architecturally sound approach - the profile sits at the highest-priority position in the model's context and shapes every response without needing to be restated. Anthropic's prompt engineering documentation covers the mechanics of system prompt design in detail, and OpenAI's prompt engineering guide documents similar best practices for their API.

If you're building a reusable prompt template that incorporates your Communication Profile alongside task-specific instructions, assembling the components in a structured editor saves considerable iteration time.

The Anti-AI Safeguard Layer

A Communication Profile tells the model what to do. You also need to tell it what not to do. Without negative constraints, the model will slip AI-isms into your voice - phrases that are statistically common in AI output but that no human writes naturally.

The difference is stark. Here's the same follow-up email, before and after applying a Communication Profile with anti-AI constraints:

Before (raw AI output):

I hope this email finds you well. I wanted to reach out regarding the Q3 proposal. Please don't hesitate to let me know if you have any questions. I'd be happy to discuss further at your earliest convenience.

After (Communication Profile applied):

Sarah - circling back on the Q3 proposal. Two things need your sign-off before Friday: the revised timeline on page 4 and the contractor budget in Appendix B. Ping me if either looks off.

The first reads like every AI email ever generated. The second reads like a specific human with a specific communication style. The Communication Profile - combined with the anti-AI blocklist below - is what bridges that gap.

Add an explicit blocklist to your profile:

ANTI-AI CONSTRAINTS:
Do NOT use these phrases under any circumstances:
- "I hope this email finds you well"
- "I wanted to reach out"
- "Please don't hesitate to"
- "I'd be happy to"
- "Thank you for your understanding"
- Any sentence starting with "I just wanted to..."

Also block structural patterns: the three-paragraph email with a pleasantry opener, a middle paragraph of substance, and a closing pleasantry. If that's not how you actually write, tell the model it's not allowed.

This is where few-shot examples become more valuable than instructions alone. You can describe your style all day, but showing the model two actual emails you've written - alongside the Communication Profile - constrains the output across dimensions that descriptions miss. The profile handles the explicit parameters. The examples handle the implicit ones: rhythm, cadence, the way you break paragraphs mid-thought.

The Self-Correction Loop

Even with a solid Communication Profile, the first output will rarely be perfect. Build a self-correction step directly into your prompt:

After drafting the email, review it against the writing samples. If any sentence sounds too polished, too generic, or uses vocabulary not present in the samples, rewrite that sentence to match the natural human patterns observed in the profile.

This instruction exploits the model's ability to critique its own output. The first pass is the generation. The second pass is a filter that catches the remaining AI artifacts. In my testing, this single addition reduces "AI-sounding" phrasing by roughly 60–70% compared to generation without self-correction.

Putting It All Together: The Complete Workflow

Here's the end-to-end process:

Step 1: Collect. Gather 10–15 writing samples. Prioritize emails, Slack messages, or any writing that represents your natural voice. Avoid polished blog posts or formal reports - those are your edited voice, not your real one.

Step 2: Extract. Run the extraction prompt above. Save the resulting Communication Profile as a standalone file: [YourName]_Style_Guide.md.

Step 3: Validate. Ask the model to write a test email using the profile. Compare it against a real email you've written on a similar topic. If it's off, identify which dimension is wrong (too formal? wrong greeting? missing your persuasion pattern?) and refine the profile.

Step 4: Persist. Store the profile where you'll actually use it - a ChatGPT Project, a Claude Project, a Gemini Gem, or a file you paste manually. Once you've found the right persistence method, every future writing task inherits the voice automatically.

Step 5: Maintain. Your writing style evolves. Every 3–6 months, re-extract from fresh samples and update the profile. Treat it like any other configuration file - version it, date it, keep the old versions.

For ongoing management, once your Communication Profile is finalized and validated, storing it in a prompt manager keeps it organized alongside your other reusable templates.

What "Good Enough" Looks Like

A well-built Communication Profile doesn't produce output that's identical to your writing. It produces output that's within the range of your natural variation. Your own emails on different days, to different people, already vary. The goal is for the AI-generated version to fall inside that variation band.