Marketing & visibility

The practice brain: teach AI to write in your voice

Published July 3, 2026 · Updated July 3, 2026

TL;DR: The reason AI drafts sound like everyone else’s is not the tool — it’s that the tool knows nothing about you. Every session starts from zero, so it defaults to the average of the internet: competent, bland, interchangeable. The fix is a practice brain: one standing document that holds your public identity, your services in your own words, your audience, your voice rules, and your hard boundaries — installed once in a workspace your AI tool reads on every conversation. Build it in about an hour (there’s a prompt below that makes AI interview you and draft it), install it, and every draft starts eighty percent done. This is the mechanic the rest of this library builds on.

Who this is for: any clinician whose AI drafts are usable but not theirs — right structure, wrong voice, and a growing pile of “make it sound more like me” edits. You’ll need one drafting engine (the tool map — Claude is the strongest fit for this particular job) with its settings done per the Safe Setup Checklist.

If you read the previous guide, you’ve met the training-wheels version: the seed document’s voice sample. This guide replaces that one paragraph with a standing file — build it once, and every workflow in this library gets better at the same time.

Why every AI draft sounds the same

Guide one made the core point: a drafting engine generates language by pattern, and it patterns on what it’s given. Give it nothing but “write a post about iron deficiency” and it gives you back the internet’s average post about iron deficiency — the same reassuring cadence, the same stock phrases, the same voice as every other practice using the same tool with the same empty prompt.

Most people fight this per conversation: paste an example, add “keep it professional but warm,” edit the output, repeat tomorrow. That works, and it’s exhausting — you’re re-teaching the tool who you are every single session, and the teaching evaporates when the chat ends.

The structural fix is the same one the content workflow applied to facts: move the expensive thing upstream and do it once. There, the expensive thing was verification, so the facts went into a seed. Here, the expensive thing is context — who you are, how you sound, where your lines sit — so the context goes into a standing document the tool reads before every draft.

What goes in a practice brain — five sections

The practice brain is a single document, one to three pages. Longer is not better: the goal is the smallest file that makes a stranger — human or machine — able to write plausibly as your practice. Five sections do it:

1. Who we are. The public identity: practice name, your name and credentials as they appear publicly, what kind of practice this is, region served. Everything here is the kind of thing already on your website’s About page — that’s the test for what belongs.

2. What we offer. Your services in your own words — not a brochure, a plain-language list of what you actually do and for whom. This is what stops the tool inventing services you don’t offer or describing the ones you do in someone else’s framing.

3. Who we’re writing for. Your audience, described the way you’d describe them to a locum: who they are, what they’re dealing with when they find you, what they’ve usually tried already, what reading level your public content targets.

4. How we sound. The heart of the file, in three parts: a short list of voice rules (“plain English first, technical term in parentheses,” “explain the why,” “no exclamation marks”), a list of words and phrasings you actually use — and a list you never use (“miracle,” “cure,” “detox,” “babe”) — and two or three paragraphs of your own published writing, pasted in verbatim as the reference sample. The rules tell the tool what to do; the sample shows it.

5. Hard boundaries. The lines that hold no matter what the draft is: education not treatment; no doses, no product recommendations, no individualized advice; no outcome guarantees or superiority claims (your college’s advertising standards, stated as standing rules); the exact educational disclaimer that goes on public content, pasted in full so the tool appends it verbatim. Boundaries in the brain get applied by default instead of remembered at midnight.

And what never goes in

The practice brain holds your general education content and brand — never patient records. No patient stories, even anonymized ones; no schedules, financials, logins, or internal disputes. Two tests, and the file passes only if it passes both: “Would I be comfortable with this document on my website?” and the PHI line from guide one. A practice brain is still a document sitting in a consumer AI tool — treat it exactly like the public-facing material it should entirely consist of.

Build it by being interviewed

You could fill in the template from a blank page, but there’s a faster way that also happens to demonstrate the whole point of drafting engines: make the tool interview you. Answering questions is easier than composing sections, and correcting a draft is easier than writing one — that asymmetry is the entire reason this tool category is useful.

The builder prompt below instructs the tool to ask you a short series of questions — one at a time, about ten minutes of answering — and then assemble your answers into the template structure, using your words rather than improving on them. You review the result the way you’d review any AI draft: what’s wrong, what’s missing, what doesn’t sound like me. Fix those, paste in your writing sample by hand (the tool shouldn’t write your voice reference — that defeats the object), and you have a v1.

Install it where the tool can always see it

A practice brain in a folder is a good start; a practice brain the tool reads automatically is the actual mechanic. That’s what standing workspaces are for.

In Claude (the lead recommendation): create a Project — a self-contained workspace with its own knowledge base and chat history. Upload the practice brain to the project’s knowledge, and put a one-line pointer in the project’s instructions (“Follow the attached practice brain for voice, services, and boundaries in everything drafted here”). Every conversation inside that project now works from the file without you pasting anything. Projects are on every plan including free (free accounts get up to five), which is more than enough for one “Practice content” workspace.

In ChatGPT: the same shape exists under the same name — Projects, on all plans: upload the file, add the pointer as project instructions (which override your general custom instructions inside the project). For lighter-weight use, ChatGPT’s account-level custom instructions can hold a condensed version of your voice rules — useful, but the two text fields are capped at 1,500 characters each, so the full brain belongs in a project.

Anywhere else: the practice brain is deliberately a plain document, so the fallback is universal — make pasting it the first message of any drafting session, in any tool. Less elegant, same effect.

One habit completes the install: the brain is a living file. Date it, keep one canonical copy in your practice docs, and re-read it when something material changes — a new service, a moved location, a voice rule you keep having to apply by hand in review. If you set a quarterly settings check from the Safe Setup Checklist, review the brain in the same sitting.

What it changes — and what it deliberately doesn’t

With the brain installed, the difference shows up in the first draft: it opens the way you open, describes your services the way you describe them, stays inside your boundaries, and ends with your disclaimer — before you’ve said anything beyond the actual request. That’s the “eighty percent done” this library keeps promising: not eighty percent of the thinking, eighty percent of the re-teaching you were doing every session.

What it doesn’t change: the brain teaches voice, not facts. A tool that sounds exactly like you can still be confidently wrong in your voice — which is arguably more dangerous, not less. The seed workflow still supplies and verifies the facts; review is still yours; the clinician still signs everything. The two documents compose cleanly, and that’s by design: the brain is the standing who and how, the seed is each week’s what. Together they’re the difference between “AI content” and your content that AI happened to format.

Install this: the template + the builder prompt

Two pieces. Run the prompt first (ten minutes of answering questions), or skip it and fill the template directly — either way, the finished file goes into your standing workspace, and your real writing sample goes in by hand.

# [PRACTICE NAME] — Practice Brain (v1 · [DATE] · review quarterly)
Companion to our AI Use Policy, tool map, and weekly content workflow.
Contains PUBLIC information only — no patient information, ever.

## 1. Who we are
- Practice + clinician(s), credentials as written publicly: [ ... ]
- Kind of practice, one sentence: [ ... ]
- Region served / where care actually happens: [ ... ]

## 2. What we offer
- [Service 1 — one plain-English sentence, in our words]
- [Service 2 — ...]
- What we DON'T offer (things we're commonly assumed to): [ ... ]

## 3. Who we're writing for
- Who they are + what they're dealing with when they find us: [ ... ]
- What they've usually already tried: [ ... ]
- Reading level for public content: [e.g., grade 8]

## 4. How we sound
Voice rules:
- [e.g., Plain English first, technical term in parentheses on first use]
- [e.g., Explain the why in 1–2 sentences when making a point]
- [e.g., Calm and specific; low adjective count; no hype]
Words/phrasings we use: [ ... ]
Words/phrasings we NEVER use: [e.g., miracle, cure, guaranteed, detox, babe]
Voice reference (paste 2–3 paragraphs of OUR OWN published writing):
[PASTE — do not let the AI write this section]

## 5. Hard boundaries (apply to every draft, unprompted)
- General education only — never diagnosis, treatment, or individualized
  advice; individual decisions are routed to a consultation.
- No doses, no product recommendations, no protocols for the reader.
- No outcome guarantees, no "better than" claims, no testimonials implying
  assured results (college advertising standards).
- Add no factual/clinical claims beyond what the request supplies;
  write [NEEDS SOURCE] where a fact is missing.
- End every public piece with, verbatim: "[PASTE DISCLAIMER]"

## 6. Housekeeping
Installed in: [e.g., Claude Project "Practice content"]  ·  Canonical copy
lives at: [ ... ]  ·  Last reviewed: [DATE]

The builder prompt — paste this into your drafting engine to be interviewed:

I’m a [PROFESSION] building a “practice brain”: one document that will teach you (and any AI tool) to draft public, educational content in my practice’s voice. Interview me to build it. Ask me one question at a time — about ten questions total — covering: who the practice is publicly; the services in my own words; who my audience is and what they’ve tried before finding me; my voice rules and the words I do and don’t use; and my hard boundaries for public health content. Then assemble my answers into the template structure I’ll paste below, using my wording rather than polishing it. Leave the voice-reference section empty for me to fill by hand, and leave [NEEDS INPUT] anywhere my answers were thin. Do not add services, credentials, or claims I didn’t state.

[PASTE THE TEMPLATE HERE]

The last instruction is the one to keep an eye on in review: the tool’s instinct is to improve your answers into brochure language, and the entire value of the file is that it doesn’t.

What’s next

The practice brain makes every draft sound like you; the next guide widens the lens from marketing to the rest of the desk: the five non-clinical jobs to hand to AI first — the highest-yield, lowest-risk places for a practice to actually start.

Sources

Get the Safe Setup Checklist

The ten things to change before your first real AI work session — including the exact privacy settings per tool.