Todd's telling this one.
On September 10, 2026, OpenAI put a voice on sale that listens and talks at the same time. Their page lists telephony support, and the examples they give are restaurant reservations and customer support. Yelp says in that same announcement it's already using it to answer booking and food order calls. Five cents a minute for the front-end voice layer. The model behind it bills separately, and there's no single published figure for that part.
So here's what's about to happen. For the next several months, half this industry will argue about whether you should put an AI on your phone line. Somebody will tell you the front desk is finished, somebody else will tell you it's all hype, and both of them are selling something.
That argument doesn't concern you yet. This does. Walk into almost any premium service business in South Florida and ask to see what they tell people on the phone. If anything's written down, it's the easy stuff. Prices. Hours. What we do.
Now ask for the hard ones. What can we not promise? What does the voicemail say at 9pm on a Saturday? What do you say to the guy who's already furious before you pick up? And the big one: at what point do you stop reading a prepared line and hand the phone to somebody who can actually decide something?
Nobody's written those down. Not the practice, not the firm, not the shop. And those four decide whether the caller stays on the line or hangs up and calls the next place on the list.
That's what this builds. Fifteen prompts and a Step 0 that pull what's in your head onto one page you can print and tape next to the phone. It works today, with a person answering. And the day you put a voice on that line, you hand it the same page, because a machine can only say what somebody bothered to write down.
If you ran last Monday's resource
You already have two of the things this one asks for: the list of questions your customers actually ask, and the question and answer cards you built from it. Bring both, skip Block 1, start at Block 2. If you didn't run it, nothing's lost, Block 1 builds them from scratch. Monday's resource is here: How to tell if AI actually understands your business.
What to have ready
- Your last ten real inquiries, names stripped out. Emails, texts, DMs, whatever the last ten people asked before they became customers. Most important item here, and the one people skip.
- Your service list with prices, however you keep it. Spreadsheet, PDF, note on your phone. It doesn't need to be pretty.
- Your real hours, including whether anybody actually picks up on Saturday. Real, not the ones on the website.
- What you don't do. The jobs you turn away, refer out, or can't touch under your license.
- How long you actually take to return a call. Not the goal. The truth.
Three rules before you start
Step 0
0. Load the material
I run a [TYPE OF BUSINESS] in [CITY]. I'm building a one-page
sheet of what we say on the phone when the question is a hard
one. I'll paste my material in pieces. Read it, don't
summarize it yet, and hold it for the rest of this
conversation.
Here's what I'm giving you:
1. My last ten real inquiries, names removed.
2. My service list with prices.
3. My real hours, including weekends.
4. What we do not do.
5. How long we actually take to return a call.
Rules for this whole conversation:
- If something you need is missing or unclear, ask me for it.
Do not fill the gap with a guess, an assumption, or an
industry average.
- If two things I give you contradict each other, tell me
which two clash and ask which one is right.
- Everything you write for me gets spoken out loud on a phone
call. It does not get read on a screen.
When you've read all five, reply with one line: what you
have, and what's missing.
Variables: your type of business, your city, and the five pieces of material.
What the output is for: everything after this runs against your business instead of a generic version of your industry.
Where this prompt fails: paste polished marketing copy instead of real inquiries and it has nothing true to work with. You'll get your own brochure back, rearranged.
Block 1 - What they actually ask
1. The real question list
From the ten inquiries I pasted, pull out every distinct
question a customer asked, in their words, not mine. Then:
1. Sort them by how often they actually came up in my ten,
most frequent first. Put the count next to each one.
2. Tag each with where it lands in a call: OPENING (first
thirty seconds), MIDDLE (they're considering), or CLOSE
(they're deciding, or pushing back).
3. Mark anything that appeared once as ONE-OFF and keep it in
a separate list at the bottom.
Do not add questions you think my industry gets. Only what's
in my ten. If fewer than ten distinct questions come out of
it, say so instead of padding the list.
Variables: none, it reads what you loaded in Step 0.
What the output is for: the real ranking of what people ask, which is rarely the one you'd have guessed.
Where this prompt fails: ten is a small sample. Something that shows up twice isn't proven common, it's a hint. Run it again at thirty.
2. The question behind the question
Go back through that list. Some of those aren't really the
question. Someone asks "how long does it take" when what they
need to know is whether they can go back to work Monday.
For every question in the OPENING and MIDDLE groups, give me:
- What they literally asked.
- What they're most likely trying to find out.
- The one detail in my answer that actually settles it.
Where you can't tell what's underneath, write "unclear, ask
the caller" instead of guessing. Keep every line short enough
to take in at a glance.
Variables: none.
What the output is for: the gap between answering correctly and answering the question. Most calls die in that gap.
Where this prompt fails: it's inferring intent from text, so sometimes it reads a worry into a plain question. Cross out the ones that don't sound like your callers.
Block 2 - The everyday answers, done properly
3. Price, out loud
Take my service list and split it in two:
A. Services I can quote a fixed price for on the phone, right
now, without seeing anything first.
B. Services where the price depends on an assessment.
For group A, write the spoken answer: the number, plus what's
included, under twenty-five words.
For group B, write it differently: what the price depends on,
the honest range if I gave you one, what happens next, and
how long that step takes. Under forty words. Never invent a
range I didn't give you. If I didn't give you one, write the
answer without a number and flag it for me.
Variables: your service list and your prices.
What the output is for: two ways to answer "how much", so the second one stops sounding like a dodge.
Where this prompt fails: if your prices genuinely depend on everything, group A comes back empty. That's information, not an error. It means your first call is a booking, not a quote.
4. Availability, honestly
Using my real hours and my real callback time, write the
spoken answers for:
- When can you come out / when can I be seen.
- Do you answer on weekends.
- I called yesterday and nobody called me back.
- I need this today.
Use the callback time I actually gave you, not a better one.
Each answer under thirty words, and each one ends with a next
step the caller can take, not with an apology.
Variables: your real hours, your real callback time.
What the output is for: the four answers that decide whether somebody waits for you or dials the next place.
Where this prompt fails: it writes what you told it. Say four hours when your team really takes two days and you've printed a promise on the wall that gets broken every week.
5. When we don't do that
Here's my list of what we don't do. For each item write the
spoken answer that:
1. Says no in the first sentence, clearly, no hedging.
2. Gives the reason in plain language, one clause.
3. Points somewhere useful: the kind of business that does
handle it, or the part of it we DO handle.
Under thirty-five words each. Don't name a specific company
you haven't been told about. Name the category.
Variables: your do-not list.
What the output is for: a no that doesn't cost you the referral.
Where this prompt fails: it can't know who's actually good in your area, so it stops at the category. Filling in real names is your job, and it's worth twenty minutes.
Block 3 - What you can never say
6. The never-say list
You are not a lawyer and this is not legal advice.
From everything I've given you, list what should never be
said on a phone call at this business: promises about
results, timelines, guarantees, coverage, safety, or any
outcome that depends on something we can't control.
For each one give me:
- The claim, written the way somebody would actually blurt it
out on a call under pressure.
- Why it's a problem, one sentence: is it a promise we can't
keep, a claim we can't prove, or something that belongs to
a licensed professional.
Eight maximum, riskiest first. Flag any that a licensed
professional in my field should review before this sheet goes
on a wall.
Variables: none.
What the output is for: the list that lives in your head, written down where the newest person on the phone can see it.
Where this prompt fails: This is not legal advice and it doesn't replace it. It's a starting list to take to your attorney, your carrier, or your licensing board, not a compliance review. If you're in medicine, dentistry, law, or insurance, there are rules a general AI chat doesn't know, and it won't tell you it doesn't know them.
7. What you can say instead
For every line on that never-say list, write the sentence
somebody CAN say instead. It has to:
- Be true no matter how the job turns out.
- Still be useful to the caller, not a wall of caveats.
- Sound like a person, not a disclaimer.
- Stay under twenty words.
Where the honest answer is "I don't know yet", write the
version that says that AND says who will know, and when.
Variables: none.
What the output is for: the half nobody writes. Tell your team what not to say without handing them the replacement and you get silence, or improvisation, which is what you were trying to avoid.
Where this prompt fails: a few come back stiff. Leave them. Prompt 13 fixes the sound of the whole thing at once.
8. The card
Turn the last two answers into a printable card.
- Two columns: NEVER SAY / SAY THIS INSTEAD.
- Eight rows maximum. If we have more than eight, keep the
eight most likely to come up in an average week and list
the rest underneath as overflow.
- Half a page. Short enough to read while a phone is ringing.
- Plain text. No design, no icons, no color.
Riskiest row first.
Variables: none.
What the output is for: the half page a new hire can use on day one with nobody sitting next to them.
Where this prompt fails: the eight-row cap is supposed to hurt. If everything feels essential you're writing a manual, and manuals don't get read next to a ringing phone.
Block 4 - The four calls nobody scripted
9. When the script stops working
A prepared answer stops working at some point in a call, and
whoever picked up needs to notice it while it's happening.
From my inquiries and everything else I've given you, name
the four signals maximum that mean this call has gone past
what a prepared answer can handle.
For each:
- What it sounds like, in the caller's own words.
- Why a scripted answer makes it worse from there.
Four maximum. If you have more, keep the four that would cost
me the most money to miss.
Variables: none.
What the output is for: the skill nobody trains and everybody needs, knowing when to stop reading.
Where this prompt fails: four is deliberately fewer than reality. This is a recognition tool for a live call, not a complete map.
10. The handoff
For each of those four signals, write the handoff. Two
versions of each.
VERSION A: the person who can decide is available right now.
The line that moves the call over without making the caller
repeat everything they just said.
VERSION B: that person isn't available. The line that gives a
real commitment instead: who will call back, and by when,
using the callback time I actually gave you.
Under twenty-five words each. Never write "let me transfer
you" with nothing after it. Never promise a time I didn't
give you.
Variables: who on your team can actually decide, and your real callback time.
What the output is for: the difference between a handoff and a hold.
Where this prompt fails: version B is worth exactly as much as your follow-through. If the call back doesn't happen, this line made the failure worse than a vague answer would have.
11. The after-hours message
Write the message that plays when nobody picks up.
- 60 words maximum. It gets heard once, not read twice.
- It answers the two most common questions on my list, the
ones I'd otherwise call somebody back just to answer.
- It says when a human responds, using my real callback time.
- It gives one clear path for an actual emergency, and says
plainly what counts as one in my business.
- Nobody says "your call is important to us."
Write it to be read out loud, then give me the word count.
Variables: your hours, your callback time, and what genuinely counts as an emergency for you.
What the output is for: sixty words that work every night, whether or not anybody's listening.
Where this prompt fails: if your business has no real emergency, and plenty don't, that line comes out forced. Cut it rather than invent one.
12. The angry caller
Write what we say when the caller is already angry when we
pick up.
The situation, generically: a customer feels something took
longer, cost more, or turned out differently than they
expected, and they are not calm.
Do not write about a specific customer or a specific case. I
am not giving you one and I don't want you to imagine one.
Keep it general enough to fit any version of it.
Give me:
- The first sentence. One. Before any explanation.
- What not to do in the first thirty seconds.
- Three lines that buy time honestly without promising
anything.
- The point where this stops being a phone call and becomes a
decision somebody has to make, plus the handoff line for
that moment.
Short enough to be read while somebody is still yelling.
Variables: none.
What the output is for: the call everybody handles worst, because it's the one nobody plans for.
Where this prompt fails: no script survives an angry caller intact. This gives whoever picked up a first sentence and a floor to stand on. Everything after that is a person, and it should be.
Block 5 - Make it usable
13. Say it, don't read it
Take everything we've written in this conversation and
rewrite it to be said out loud.
That means:
- Contractions everywhere. Nobody says "do not."
- Sentences a person can finish in one breath.
- No corporate words: utilize, assist, endeavor, at this
time, we apologize for any inconvenience.
- No sentence that would sound strange coming out of somebody
standing in my shop.
Read every line back to yourself before you keep it. If it
sounds like a form letter, rewrite it.
Give me only the final version. No before and after, no notes
on what you changed.
Variables: none.
What the output is for: the difference between a sheet people use and a sheet people feel strange saying out loud, which is the same as a sheet nobody uses.
Where this prompt fails: it overcorrects. Some of the careful language from prompt 7 has to stay stiff, because stiff is what makes it safe. Put those lines back.
14. Build the page
Build the final sheet. One page.
- Numbered sections, in the order a call actually goes:
opening questions, price, availability, what we don't do,
never say / say instead, the four signals and their
handoffs, after hours, angry caller.
- 60 lines maximum, total. If it runs long, cut the least
frequent answers, never the hardest ones.
- No explanations. No heading longer than four words. Answers
only.
- One line at the bottom: who to ask when the sheet doesn't
cover it.
Then separately, give me a placement line I can paste at the
top: this sheet lives printed next to the phone, and as a
file in the team's shared folder. It does not live in an
email.
Variables: who your team asks when the sheet runs out, and where your shared folder lives.
What the output is for: the thing itself. Print it.
Where this prompt fails: sixty lines is what fits on a page somebody glances at mid-call. Let it run to three pages and you've built a binder, and the binder goes in a drawer.
15. Five test calls
Write five test calls I can run on somebody from my team,
holding the sheet, today.
- Four normal ones, drawn from the most frequent questions.
- At least one hard one: an angry caller, a question that
should trigger a handoff, or something off the never-say
list.
- For each, what the caller says first, plus one follow-up to
use if the answer comes back vague.
- For each, tell me which numbered section of the sheet holds
the answer.
If any of the five has no section on the sheet, say so
plainly. That gap is the whole point of the exercise.
Variables: none.
What the output is for: the only real test, somebody using the sheet under a little pressure while you watch.
Where this prompt fails: a colleague in the room is not a stranger on a bad connection with a kid crying behind him. Run the five, then listen to real calls for a week and fix what the test missed.
The six pieces every one of these prompts has
Any prompt above has the same six parts. Once you can see them, you can write your own for anything, not just the phone.
| Piece | What it does | Where it shows up here |
|---|---|---|
| Channel | Says where this lands, so the writing fits the place | "Everything you write for me gets spoken out loud on a phone call" (Step 0) |
| Input | Names the raw material and refuses to work without it | "My last ten real inquiries, names removed" (Step 0) |
| Cap | A hard limit that forces a choice instead of a pile | "60 words maximum" (11), "Eight rows maximum" (8) |
| Ban | What it must not do, stated before it does it | "Never invent a range I didn't give you" (3) |
| Shape | The exact form the answer comes back in | "Two columns: NEVER SAY / SAY THIS INSTEAD" (8) |
| Escape hatch | What to do when it can't, so it doesn't make something up | "Write 'unclear, ask the caller' instead of guessing" (2) |
The one specific to this resource is the channel. Every instruction here tells the AI it's writing for speaking, not for reading. Drop that line and you get prose. Correct prose, sometimes good prose, and useless to a person holding a phone.
Where all this falls short
- It installs nothing. These prompts produce a sheet of paper. Putting a real voice on your phone line needs a telephony provider on top of the model, plus somebody to wire it up and keep it running.
- The sheet doesn't apply itself. Printed and taped up is step one. People using it under pressure is step two, and step two is the hard one.
- Block 3 is not legal advice. It's a starting list for a conversation with somebody licensed to have it with you.
- Skip the ten real inquiries and the whole thing goes generic. Every prompt leans on that material. Without it you get a competent script for a business that resembles yours from a distance.
- We have not run this with an automated voice, and we're not claiming it works there. What we have done is write answer sheets for people who answer phones. That the same page feeds a voice system later is a reasonable argument, not a result.
- The five cents is the voice layer only. The model behind it bills separately and there's no published total. Anybody quoting you a full per-minute cost today is estimating.
Keep reading
- How to tell if AI actually understands your business - produces the question list and answer cards this resource assumes you have
- Cited or bought - the two ways to show up inside the answer your customer reads
- A ChatGPT citation won't pay your bills - what to do when they find you and the phone still doesn't ring
- Where your client data lives - why this sheet has to live somewhere that belongs to you
- SEO vs. GEO vs. AEO: which one does your business need - the five disciplines, and which ones your business actually has to care about
