This is the file that turns a blank worker into a specific agent. Fill it in, paste it into
your agent's JOB setting, and that is now what it does.
Everything below has a reason to exist. Skip a section and your agent gets vaguer, which
shows up later as output you have to rewrite.
Nothing in here is technical. You are describing a role to a very capable assistant who
has never met you, starts every shift with no memory, and will do exactly what you wrote
rather than what you meant.
Staring at empty fields is the slow way to do this. Copy the whole block below and paste it
into a new chat with Claude or ChatGPT. It will ask you about your practice, write the job
description for you, and hand it back ready to paste.
You are helping me write a job description for an AI agent I am about to build. I run a
health and wellness practice. I am not technical, and I have never done this before.
Interview me. Do not lecture me and do not hand me a form.
Run it like this.
find me, and what I offer. Ask only what THIS agent's job needs. Do not interview me
about my whole business.
is faster for me than inventing an answer from nothing. Offer me your best guess in a
sentence and ask whether it is right.
nowhere on your list.
Two things you must push back on, because I will get them wrong.
the format, the length, and the order. Ask me what I would want to see on screen.
cannot be undone**, tell me to make that a draft I approve instead, and tell me why.
Being wrong there costs me a relationship.
When you have enough, write the finished job description under these seven headings.
1. Who it is, and what it owns (including what it does NOT touch)
2. What it does, every time it runs, as numbered steps
3. What it hands me, with the format, length and order stated precisely
4. How it should sound (only if it writes things I send)
5. When it stops and asks me (decides alone, drafts for me, or stops)
6. What it needs to know about my practice
7. When it runs
Write it addressed TO the agent, as instructions it will follow. Put the whole thing in
one code block so I can copy it in a single go. Then tell me the one part you are least
sure about, so I know where to look first when I see its output.
Start now, with your first question.
*(Answer as if you were briefing a new assistant on their first morning. Ordinary language
beats professional language here. The parts where you sound most like yourself are the parts
that make the agent's output sound like you.)*
Give the role one line, then draw the boundary, which does more work than the title does.
Role. _______________________________________________
Owns. _______________________________________________
Does not touch. ______________________________________
Why the boundary. An agent with no edges wanders. It will helpfully rewrite something you
never asked it to touch. Naming what it does NOT own is what keeps its output narrow enough
to trust.
Example
Role. My follow-up assistant.
Owns. Anyone who inquired and never got a reply, and anyone who went silent after
showing interest.
Does not touch. Existing clients, anything about their care, anything about money.
The actual job, as steps. Be specific enough that two different people reading it would
produce the same thing.
1. _______________________________________________
2. _______________________________________________
3. _______________________________________________
Why steps. "Handle my follow-ups" is a wish. "Find X, check Y, draft Z" is a job. The
difference between them is whether you can tell when it did the work wrong.
Example
1. Go through the inquiries you are given and find anyone who reached out and never got a reply from me.
2. Find anyone who showed interest, heard back from me, and has gone silent for more than a week since.
3. For each person, draft a short message I could send, picking up from what they asked about.
Describe the output precisely, meaning the format, the length, and the order. This is the
section people skip, and it is the one that decides whether the agent saves you time or
creates a second job of reformatting.
_______________________________________________
The test. Can you look at what it produced and act on it in under a minute? If checking
its work takes as long as doing the work, the problem is in this section, so come back and
make it sharper.
Example
A short list. One person per line, their name, when they first reached out, one line on
what they asked about, then the draft message underneath. Newest first. No introduction,
no summary, no encouragement.
This only comes into play if it writes anything you send. When it does, fill this in
properly, because the fastest way to lose someone is a message that sounds like it came from
a company.
Sounds like. ________________________________________
Never says. _________________________________________
Example
Sounds like. A person who remembers them. Warm, short, no pressure, plain words.
Never says. "Circling back," "just checking in," "touching base," anything that sounds
like a sales sequence.
This is the section that decides whether you end up with an agent you trust or one you
double-check forever.
It decides alone. ____________________________________
It drafts, you send. _________________________________
It stops and asks. ___________________________________
The rule that works. Anything reversible and private, it can do on its own. Anything that
reaches another human, or moves money, or cannot be undone, it prepares and you approve.
Start stricter than feels necessary. Loosening this later takes ten seconds. Discovering it
sent something you would never have sent is a different kind of day.
Example
It decides alone. Who makes the list, and the order they appear in.
It drafts, you send. Every message. Nothing reaches anyone until I have read it and sent it myself.
It stops and asks. Anyone who sounds upset, asks about a refund, or turns out to be an existing client. It lists them for me with no draft.
Your agent starts every run knowing nothing. Give it the context THIS job needs, which is a
much smaller thing than your whole business. Just the part this role actually touches.
_______________________________________________
Example
I work with people in perimenopause who have usually been told their labs are normal. They
have often been researching for a year before they find me. Most inquiries come through
the contact form or Instagram. My first session is 90 minutes and costs $240.
Schedule. ____________________________________________
Pick a time you will actually look. An agent that runs at 3am and produces something you
see on Thursday is a filing cabinet. One that runs an hour before you start is a colleague.
Example
Schedule. Every weekday at 7am my time, so the drafts are waiting before my first client.
Sections 1 through 7, one after another, becomes the JOB setting on your agent worker. That
is the whole handover. What you are writing is a role, in ordinary words, and the role is the
part only you can write.
Come back to this file. The first version of any agent is wrong in small ways you will
only see in its output. Change one section, run it again, look. That loop is the skill, and
it is the reason you build these yourself instead of being handed someone else's.
*Yours to change entirely. If a section does not fit the agent you are building, cut it. The
structure is here to stop you forgetting something that costs you later, rather than to be
obeyed.*
Post it in the VIP chat and ask your question there. If a step does not work, if your screen does not match what is written here, or if something just reads confusingly, that is worth saying out loud.
I am building this while you are walking through it. If you do not tell me a thing is broken, I do not know it is broken, and it stays broken for whoever comes next. Tell me and I will fix it as fast as I can.
Wins go in the same place. They tell me which parts are landing.