Don't Let AI Do Your Thinking: What to Tell It Before It Writes

"Do not let AI outsource your thinking" has become a common line. It's a good line. What's usually missing is the how.

Here's the version I work by. I've been studying AI since 2021 and building with it since 2023, and the pattern is the same every time: the brief is the thinking. AI is a tool. It writes fast and it writes fluently. But it can't know who your clients are, what you're allowed to say, or what good looks like in your practice. Those are decisions. They belong to you.

The same request, done twice

Say a practice manager at an allied health clinic needs an email to families about a change to cancellation fees.

The lazy prompt:

What comes back is polite, generic and slightly wrong. It may invent a notice period. It might sound like a bank. It might promise flexibility you don't offer.

The briefed prompt tells the tool who the email is for, what has changed and why, the exact policy wording, what tone fits these families, what it must not say, and who will check it before it goes out.

What comes back from that is something a person can review in two minutes and send. The difference isn't the tool. It's the five decisions made before typing.

The five decisions only you can make

1. Who is it for?

Not "our clients". Which ones? Parents of young children on NDIS plans read differently from adult private clients or support coordinators. Name the reader and what they're worried about.

2. What does good look like?

Decide the outcome before the words. "Families understand the change, know when it starts, and know who to call with questions." If you can't say what good looks like, the AI can't either.

3. What facts may it use?

Paste in the real facts: the policy wording, dates, contact details. Then tell it to use only those. AI fills gaps with plausible guesses, so remove the gaps.

4. What must it never say?

In healthcare this list matters. No clinical advice. No promises about outcomes. No pricing you haven't confirmed. No client details. No superlatives. Write your own list once and reuse it.

5. Who checks it?

Name the person who reads it before it goes out, and what they're reading for. If nobody owns the check, the AI is making the final decision by default.

Turn the decisions into a reusable brief

Once you've made these decisions for one type of message, save them. A short brief template means your team doesn't start from a blank box each time.

TASK: Write [type of message, e.g. an email to families].

WHO IT'S FOR: [specific reader, e.g. parents of children on NDIS plans who
attend weekly sessions]. What they're likely worried about: [e.g. being
charged unfairly, losing their regular time slot].

WHAT GOOD LOOKS LIKE: After reading, they [e.g. understand the change, know
the start date, and know who to contact].

FACTS YOU MAY USE (use only these; write [CHECK] if you need anything else):
- [fact 1, e.g. exact policy wording]
- [fact 2, e.g. start date]
- [fact 3, e.g. contact name and phone]

TONE: [e.g. warm, plain, respectful, Australian spelling, short paragraphs].

NEVER: give clinical advice, promise outcomes, mention prices not listed
above, include any client details, use words like "best" or "leading".

LENGTH: [e.g. under 180 words].

Before the draft, list any questions you'd need answered to do this well.
After the draft, list every sentence you're not sure is accurate.

The last two lines matter most. Asking the tool to list its questions first shows you what you forgot to decide. Asking it to flag its unsure sentences tells your reviewer where to look.

The check step

Before anything goes out, a person reads for five things:

• Accuracy. Every fact matches the source. Nothing invented.
• Scope. Nothing clinical, nothing that reads as a promise.
• Privacy. No client names or identifying details, in the draft or in what was pasted into the tool.
• Voice. It sounds like your practice, not a call centre.
• Action. The reader knows what to do next.

If the draft fails any of these, fix the brief, not just the draft. That way the next one comes out better.

Training a team to brief, not just "use AI"

Telling staff to "use AI" usually means everyone types a one-line request and edits the result for longer than it would have taken to write. A better start:

1. Pick the three messages your team writes most often.
2. Fill in the brief template for each, together, as a team.
3. Save the briefs somewhere shared and name an owner for each.
4. Agree on the check step and who does it.
5. Review the briefs every few months as policies change.

This keeps the thinking inside your practice, where it belongs.

What a person must still decide

• Who the message is for and what it should achieve.
• Which facts are true, current and allowed to be shared.
• What your practice will never say.
• Whether the final draft goes out, gets fixed or gets binned.
• When a message is too sensitive to draft with AI at all.

Your checklist

• Name the reader and what they're worried about
• Write what good looks like in one sentence
• Paste in only the real facts
• Write your "never say" list
• Name the person who checks it
• Ask the tool for its questions first and its unsure sentences last
• Save the brief as a template for next time
• Use made-up example details when practising, never real client information

If you'd like this set up for you, that's what the free check is for.

General information only — not clinical, legal or financial advice. No client is named and nothing here describes a patient or participant. Written by Aishah Shah, Western Sydney.