What the model developers themselves advise
The three companies whose models small businesses use most have public guides to writing prompts: OpenAI, Anthropic and Google. They are written differently, but they agree on the same parts of an instruction.
✓ — the guide gives direct advice; ≈ — the topic is mentioned more generally. Guides checked on 12 October 2026.
Two details from these guides are worth remembering. Anthropic suggests picturing the model as a capable but new employee who does not know your rules and routines: the more precisely you explain, the better the result. Google recommends always including examples and says prompts without them are likely to be less effective.
What a ready-made prompt lacks
Now put a prompt from a collection next to that list. The collection's author can write a role ("you are an experienced sales manager") and a format ("under 100 words, polite"). The rest they cannot write, because they do not know your business:
- context — your delivery terms, prices, return rules, what you may promise a customer and what only a manager decides;
- examples — your answers that worked, and your tone;
- testing — what counts as a mistake for you: a promised discount, a wrong deadline, a mention of a competitor.
Without context the model fills the gaps with something plausible. For a letter to a customer that means an invented delivery date or a compensation you do not offer. One of the models made exactly this mistake in our AI assistant test, on the task of writing to a customer: it added facts and compensation that were not in the brief. Details are in the article on the AI assistant test.
There is a second reason. OpenAI says plainly that different model types, and even different versions of the same model, can respond differently to the same prompt. A prompt the collection's author tested on last year's model may behave differently on the model you use today.
Before and after on one task
An illustrative example: an online shop selling outdoor gear answers complaints about late parcels. Here is a typical prompt from a collection.
You are an experienced customer support manager. Write a polite
and professional reply to a customer complaining about a late delivery.
It has a role and a general wish about tone. The model does not know when the parcel was sent, where it is now, or what the shop can offer. So it will either write generic apologies or invent details.
Here is the same task written in five parts. The shop's rules are illustrative.
ROLE
You reply to customers on behalf of an online shop selling outdoor
gear. You write in Ukrainian, politely, briefly, without officialese.
CONTEXT
- We ship with Nova Poshta within one working day after payment.
- If a parcel has been in transit for more than 4 days, a manager
files a trace request.
- Only the owner decides on compensation or discounts for delays,
so do not promise them in the reply.
- Order data: dispatch date, waybill status.
FORMAT
Up to 5 sentences. First, acknowledge the problem. Then say what
the waybill status shows and what we are doing next.
Do not give dates or deadlines that are not in the data.
EXAMPLES
Two or three of our replies to similar complaints that customers
rated well.
CHECK
Before returning the reply, check: no promise of compensation,
every date comes from the order data, the next step is named.
CUSTOMER COMPLAINT
...
Each added line closes a gap where the model would otherwise guess: dispatch time, the trace rule, the limit of authority, the length of the reply. The examples set the tone better than the word "polite". The check section spells out what counts as a mistake for you.
You do not need to pass the customer's phone number, full name or address in such requests: the dispatch date and waybill status are enough for the reply. Which data you can give AI services is covered in the article on data for an AI assistant.
The instruction template
| Part | What to write | Where to get it |
|---|---|---|
| Role | On whose behalf, for whom, in what tone | One or two sentences, written in a minute |
| Context | Facts and rules without which the answer will be wrong; what must not be promised | Your delivery, payment and return terms, price list, procedures |
| Format | Length, structure, language, what comes first | How you want the result: a letter, a table, a list |
| Examples | 2–5 good results for similar cases, different from each other | Your best replies, descriptions, letters from recent months |
| Check | A list of mistakes that must not happen | Your real complaints and corrections: what has already gone wrong |
The developers differ only in formatting details. Anthropic suggests separating the parts with XML tags such as <context>, OpenAI with Markdown headings or XML. For work in the ChatGPT, Claude or Gemini chat, plain capitalised headings like those in the example above are enough.
Testing on real cases
An instruction you will use every week is worth testing before the whole team starts using it.
- Collect 10–20 real cases — complaints, requests, products. Pick varied ones, awkward ones included.
- Write down the criteria before you start: what the answer must contain and what it must not. The same points as in the instruction's check section.
- Run all the cases and mark each answer: usable as is, needs editing, unusable.
- Fix the instruction, not the answers. Every unusable answer shows what is missing from the context or the examples. Add it and run the same cases again.
- Repeat the test after a model change. OpenAI explains why: a new model version can behave differently on the same prompt.
The same approach, on 50 questions, is described in the article on choosing a chatbot. A bot's instruction is the same kind of prompt, only longer.
Ready-made prompt collections have one useful purpose: seeing which tasks you can hand to AI at all. They are not worth using as finished instructions. Which tasks are better not handed over is covered in the article on AI for business.
Our position
For every task that recurs weekly, write one five-part instruction and keep it as a company document, next to your procedures. Take the examples from your own work. Test the instruction on real cases before rollout and every time you change the model or plan.
This does not hold for one-off tasks: a single question, a single text you will read and correct anyway. There a short request in your own words is faster than a five-part instruction, and you will check the result straight away.
The developers' advice was checked on 12 October 2026 on the prompt-writing guide pages of OpenAI, Anthropic and Google. The example instruction is illustrative; the shop's rules in it were made up for the illustration.