A prompt library is only worth something if it knows how AI fails on the jobs you actually do. This is the list: the specific mistakes these prompts are written against, how each one shows up in a draft, the guard inside the prompt, and the check that catches it before a customer does.
Nine failure modesEvery one has a guardLast updated 24 October 2026
What this page is, and what it is not
It is a working list of the failure modes the library is written against. Each entry names the mistake, shows the shape it takes in a draft, and says what the prompt does to prevent it and what a person still has to check.
It is not a set of customer cases. The draft lines on this page are examples of the shape, written for this page — not messages from a real customer, and nothing here identifies a person or a business.
It is also not a claim that the library cannot fail. It can. That is why every kit ends with checks, and why the free checker exists: the last line of defence is a person reading the draft against the facts.
Failure 01
It invents a price
The draft reads like a quote because it has a number in it. Nobody supplied that number.
What it looks like
“Our call-out is $85, and the part is usually around $40, so about $125 all in.” No price was given to the tool.
Why it happens
A model asked for a quote reply will produce the shape of a quote, and the shape includes a figure. It has read thousands of price lists; it has not read yours.
The guard
Every quote prompt says to use only the supplied figures, to leave anything missing as a visible bracket, and to finish with an “assumptions to confirm” list. The quote kit starts with the numbers before it writes a word.
Your check
For every figure, ask where it came from. If the answer is “the AI worked it out”, it is not a fact. The free checker at /tools/check flags prices that are not in the facts you paste, and the quote calculator shows the range and the margin behind it.
Failure 02
It gives something away
An apology turns into a free meal, a refund or a discount that nobody approved.
What it looks like
“Please come back in for a free meal on us, and we'll refund the last one.” Two promises, made in one sentence, by a tool that does not know your margin.
Why it happens
“Make it right” is an instruction a model resolves by offering something. The offer costs the model nothing.
The guard
The prompts say no free item, refund, credit or compensation unless the owner has approved that exact offer. The review builder leaves the goodwill decision to you and fills in only what you supplied.
Your check
Search the draft for free, refund, discount, credit, on us, no charge. Each one needs a person's name against it before it goes out. The checker flags all of them.
Failure 03
It invents a date or a time
“Thursday at 9” sounds helpful and is the most expensive phrase in a small business.
What it looks like
“We can get to you Thursday morning, around 9.” The diary was never mentioned.
Why it happens
A reply about a booking expects a time. The model produces a plausible one, and plausible is not booked.
The guard
Every scheduling and quote prompt asks for the real slots, and holds the sentence open until they are supplied. The kit for a booking starts by listing the diary, not by writing the message.
Your check
Check every day and time against the diary, and the year against the date. If the diary has not been opened, the message is not ready. The checker flags dates and times with no source.
Failure 04
It names a staff member
A friendly detail in a public reply, and a person's name attached to a complaint they did not agree to.
What it looks like
“I've spoken to Maria, who served you, and she's sorry.” Maria was not asked, and the reply is public.
Why it happens
Naming someone reads as accountability. In a public thread it is also a disclosure.
The guard
The review prompts say no staff member's name in a public reply, and the reply builder writes “the team” unless you supply a name on purpose. Anything about a named person is routed to the owner first.
Your check
Read the draft as the staff member. Would they be comfortable seeing it on the internet, next to their name? The checker flags supplied staff names and role-plus-name patterns.
Failure 05
It admits fault
An apology that accepts a legal conclusion you have not agreed to.
What it looks like
“We damaged the pipe when we were there, and we're liable for the ceiling.” Nobody has established what happened yet.
Why it happens
Apologising is the polite move, and a model does not distinguish between “sorry this happened” and “we are legally responsible”.
The guard
Prompts keep the apology and remove the legal conclusion, and any message touching damage, injury or a dispute is written as a holding reply until the owner has the facts.
Your check
Look for fault, liable, damaged, should have, failed to. Keep the human sentence; take out the verdict. The checker flags admissions and passes the wording to you.
Failure 06
It leaves a bracket in
The draft is good except for [YOUR NAME], which goes out to a customer.
The prompt asked the model to leave missing facts visible, which is the right behaviour. The mistake is skipping the last read.
The guard
Brackets are deliberately loud: uppercase, square, impossible to skim past. The kit's checklist says fill or delete every one, and the saved profile fills the common ones automatically.
Your check
Search the draft for the [ character before sending. The checker lists every bracket it finds, which is the fastest version of the same check.
Failure 07
It invents a credential or a claim
“Fully certified, 24/7, rated five stars” — three claims, none of them yours.
What it looks like
“We're fully certified, available 24/7, and rated five stars by our customers.”
Why it happens
Marketing copy is full of these phrases, and a model writing an advert reaches for them. They are also the claims a regulator or a competitor will ask you to prove.
The guard
Prompts say no claim that cannot be evidenced, and the listing and advert kits ask for the qualification, the hours and the actual review count before writing.
Your check
For each claim, name the certificate, the rota or the number. If you cannot, cut it. The checker flags certification, availability and rating language for review.
Failure 08
It invents availability
A slot that does not exist, offered in writing, to a customer who now expects it.
What it looks like
“We have space tomorrow afternoon and can fit you in this week.” The week is already full.
Why it happens
“Can you fit me in?” invites a yes. The model has no calendar and no reason to say no.
The guard
Availability prompts require the real openings, and the booking kit separates “holding reply” from “confirmed slot”. A holding reply commits to nothing but a callback.
Your check
Read the diary before the message goes out, and prefer the holding reply when you cannot check. A callback promise is cheap; a phantom slot is not.
Failure 09
It invents a policy
A generous returns line, a guarantee, or a cancellation term that your business does not actually offer.
What it looks like
“Of course — our no-quibble 30-day guarantee covers that.” There is no such guarantee.
Why it happens
Customer-service language assumes a policy exists. The model writes the friendly version of a rule you never made.
The guard
Policy and terms prompts take your existing wording as the only source, mark anything missing, and say where the law or a scheme overrides the draft. The free policy builder does the same for an AI use rule.
Your check
Compare the sentence with the policy you can point to. If it is not written down anywhere, it is not a policy yet — it is a promise made in a message.
Keeping this honest
How a new entry gets written
This page grows when a new failure turns up while writing or checking a kit. Every entry uses the same four fields, so it is useful rather than a story:
What it looks like. One line of draft, as close to the real shape as possible, with no customer detail in it.
Why it happens. The reason a model does it, in one sentence. If nobody can say why, it is not understood yet.
The guard. The exact instruction in the prompt or kit that prevents it, and the tool that catches it if the guard is skipped.
Your check. What a person does in ten seconds before the message leaves. If that step cannot be described simply, the guard is not finished.
If a draft fails in a way that is not on this list, that is worth knowing: send the shape of it, with the customer detail removed.
This prompt may need customer, staff, financial or confidential detail to be useful. Strip it down before it goes into an AI tool:
Names and contact details. Use “the customer”, “the tenant”, “the applicant”.
Addresses, access codes, alarm codes and key-safe numbers. Replace with a place, or leave the bracket empty.
Card, bank and invoice details. Never paste them. Refer to “the outstanding balance”.
Health, safety or disciplinary detail. Keep it to the administrative shape of the case, not the person.
Anything you would not put in an email to a stranger.
Then read the draft as the person it is about. A model does not know your policies, your contract or your regulator — the check on each kit says where the judgment stays.
One step before payment
Where the receipt goes
Polar takes the payment and sends the receipt. These two fields travel with you to checkout, and the portal that comes with the subscription is where you change the card, get invoices or cancel.