Try it · how we prompt
Every free prompt on this site is built the same way: six labelled parts, then a space for your details. Once you can see the parts, you can write your own for any job in the business.
The six parts
This is the review replies prompt, exactly as it is on its own page. Each part does one job.
01
ROLE
Tell the AI who it is working for and as whom. A reply written as the owner sounds different from one written as a chatbot.
You are helping the owner of a small UK hospitality business, a café, restaurant, pub or bar, reply to a customer review. Write as the owner, using "we".
02
CONTEXT
One or two lines on the situation. It explains the rules that follow, so the AI can follow their spirit and not just their letter.
Review replies are public. Other customers read them to judge how we treat people, so a good reply is short, warm and specific, and it never argues.
03
THE JOB
The task, step by step, ending with the start rule: if your details are not there yet, ask for them and wait.
Write one reply to the review in MY DETAILS. - Thank them by referring to one specific thing they mentioned. - If they raise a problem, acknowledge it without arguing and say what we will look into. If the review is about an allergy, illness, injury or discrimination, or threatens legal action, do not draft a reply: tell me to get advice before replying in public, and stop. Start rule: if the review is in MY DETAILS, use it. If it is missing, or only the brackets are there, write nothing yet. Ask me to paste the review and anything I know about what happened, then wait for my answer.
04
FORMAT
Length, shape and language. If you want 40 to 80 words in British English, say so, or you will get 300 in American.
- British English, warm and plain. The reply says "we" throughout, never "I". - 40 to 80 words. - End with "[Your name]" on its own line so I can sign it. - The reply first, ready to post, then the checks below. Plain text: no headings, dividers or bold.
05
WHAT NOT TO DO
The mistakes that would cost you: promising a refund in writing, naming a member of staff, making something up.
- Do not offer or promise refunds, discounts, free items or compensation in writing, and do not hint at one. If they ask for something back, invite them to get in touch with us directly and say no more about it. - Never name staff or other customers, even if the review does. - Never add details that are not in the review or my notes: no times or dates, no job titles, no actions we have not told you about, and no claims about what we are usually like. - No brackets to fill in except [Your name]. - No exclamation marks, no emoji.
06
HOW TO CHECK THE ANSWER
The AI checks its own draft against the rules, then tells you what to check yourself. You still read every word.
Before you show me the reply, check it against every rule above and fix anything that breaks one. Under the reply, add a short list headed "Check before you post": what I should check myself about this review, such as whether the facts match what happened, and anything you were unsure of. Then read the reply and the list once more and delete anything that is not in the review or my notes, such as an action, a cause, a piece of equipment or a time.
Then
MY DETAILS
The brackets are where you paste your review, rota or figures. Use initials or roles, never full names.
The review: [paste the review here] What I know about what happened (optional): [paste notes here, or leave blank]
The start rule
When you tap Open in ChatGPT, Claude or Grok, the prompt arrives before your details do. Some apps send it straight away. The start rule tells the AI to ask for your details and wait, instead of answering about empty brackets. So the first thing you see is a question, and the second is your answer.
Checking the answer
How we test them
We run every prompt on three invented examples and one empty start, through GPT, Claude and Gemini: 12 answers. Each answer is checked against a written list, once by us and once by a second AI that never sees our verdict. If anything misses, we change the prompt and run all 12 again. If we publish a prompt with misses, its page says exactly how many.
The rota is the hard one. Its three misses got someone’s hours wrong, booked a double shift, or suggested hiring cover nobody mentioned. That is why the rota page tells you to check every rota yourself.
Tested 3 October 2026 on our three prompts, through each maker's own connection, which may not be the exact model a free app gives you. Each prompt gets three examples and one empty start; an answer counts only if both checkers passed it (Claude Opus 5.5 and GPT-5.5, each blind to the other).
| AI model | Reviews | Rota | Stock | Total |
|---|---|---|---|---|
| Gemini 3.5 Flash | 4 of 4 | 3 of 4 | 4 of 4 | 11 of 12 |
| Claude Sonnet 5.5 | 4 of 4 | 3 of 4 | 3 of 4 | 10 of 12 |
| GPT-5.5 (through codex, house instructions loaded) | 4 of 4 | 2 of 4 | 4 of 4 | 10 of 12 |
| Qwen 3.6 27B (free, runs on our own machine) | 4 of 4 | 2 of 4 | 4 of 4 | 10 of 12 |
| Claude Opus 5.5 | 3 of 4 | 2 of 4 | 4 of 4 | 9 of 12 |
| Gemini 3.1 Pro (preview) | 3 of 4 | 3 of 4 | 3 of 4 | 9 of 12 |
| Claude Haiku 4.5 | 2 of 4 | 1 of 4 | 1 of 4 | 4 of 12 |
| Gemini 3.1 Flash-Lite | 1 of 4 | 1 of 4 | 1 of 4 | 3 of 12 |
The biggest model was not the best. Gemini 3.5 Flash got 11 of 12 right, ahead of Claude Opus 5.5 and Gemini 3.1 Pro on 9.
The smallest, cheapest models struggled: Claude Haiku 4.5 got 4 of 12 and Gemini 3.1 Flash-Lite 3 of 12, often skipping the rota table or inventing figures.
A free model running on our own machine, Qwen 3.6, got 10 of 12, level with GPT-5.5 and Claude Sonnet 5.5.
Every model got at least one rota wrong, so whichever AI you use, check the rota yourself.
Write your own
Copy this, fill in each part for one job that eats your week, then paste it into the AI you use. Keep the start rule and the check as they are.
Before you paste
Use initials or roles, not full names, and leave out health, contact and other personal details. What you paste goes to the AI company you use, so check its privacy settings.
Or start from a tested one: review replies, the Sunday rota or stock and prices.
ROLE [Who should the AI be, and who is it working for?] CONTEXT [What is going on, and why does it matter?] THE JOB [What should it do, step by step?] Start rule: if MY DETAILS is empty, or only the brackets are there, do nothing yet. Ask me for my details, then wait for my answer. FORMAT [How long, what shape, British English.] WHAT NOT TO DO [The lines it must not cross.] - No exclamation marks, no emoji. HOW TO CHECK THE ANSWER Before you show me the answer, check it against every rule above and fix anything that breaks one. Then add a short list headed "Check before you use it": what I should check myself, and anything you were unsure of. MY DETAILS [paste them here]
Compare notes
Send us your admin week, or just the one job that hurts most. The first conversation is free.
Email hello@kap.link · how we handle your data
Email Ollie