A prompt is the message you type to an AI to tell it what you want: a question, an instruction or a task. "Summarise this letter in three bullet points" is a prompt. The AI only knows what you give it in the prompt and the conversation, so a clear prompt with the right context, a clear task and the format you want gets far better results than a vague one-liner.
An everyday analogy
Think of briefing a new temp on their first morning. If you say "sort out the post", you get something. If you say "open today's post, put invoices in the blue tray, anything from the tax office on my desk, and tell me if anything is urgent", you get exactly what you need.
An AI is that temp: capable, fast, and with no idea how your office works unless you say so.
The four parts of a good prompt
- Role or context: who the AI should act as, and the situation. "You are helping a GP practice write to patients."
- Task: exactly what to do. "Write a short letter explaining that Friday clinics are moving to Thursday."
- Details: facts, tone, limits. "Friendly, plain English, under 120 words, no medical jargon."
- Format: how you want it back. "Give me a subject line and the letter."
You do not need all four every time. But when an answer disappoints, one of them is usually missing.
Five AI prompt examples (fictional)
1. Clinic: patient information
Prompt: "You are writing for patients of a general practice. Explain in plain English, at a reading age of 12, what a fasting blood test is and how to prepare. Use five short bullet points. End with: If you have questions, call the practice."
2. Clinic: structuring notes
Prompt: "Turn these rough notes into a structured note with the headings Complaint, Findings, Assessment, Plan. Do not add anything that is not in the notes. If something is unclear, list it under To check."
(Paste only the clinical content, without the patient's name or other identifiers, unless your tool is approved for patient data.)
3. Law firm: chronology
Prompt: "Below is an email chain about a property dispute. Create a chronology as a table with columns Date, Who, What happened. Use only facts stated in the emails. Flag any dates that conflict."
4. Law firm: plain-language explanation
Prompt: "Explain the attached clause to a client with no legal background. Two short paragraphs, neutral tone, no advice on what they should do. Then list three questions the client might ask."
5. Office: meeting follow-up
Prompt: "Here are my notes from today's supplier meeting. Write a follow-up email confirming the three agreed actions, who owns each one, and the deadlines. Professional but warm. Put the actions in a numbered list."
Notice what they have in common: a role, a precise task, limits ("only facts stated", "do not add anything"), and a clear format.
Tips that make a big difference
- Give the source. Paste or upload the document rather than asking from memory. AI answers are more reliable when grounded in your material.
- Say what not to do. "Do not invent case law" or "do not add new diagnoses" reduces made-up content.
- Ask it to flag uncertainty. "If you are not sure, say so."
- Iterate. Your first prompt rarely gets the final version. Reply with "shorter", "more formal", "add the deadline".
- Start fresh for new topics. Very long chats fill the AI's working memory and quality drops: see context rot.
Prompts are measured in tokens, small pieces of text. Our guide what are tokens in AI explains why long prompts cost more.
What not to put in a prompt
Treat a prompt like an email to someone outside your organisation. Unless the tool is approved for it, leave out:
- Names, dates of birth, addresses, insurance or case numbers of patients, clients or employees.
- Health details linked to a person: these are special category data under the GDPR.
- Passwords, access codes, bank details.
- Unreleased financial figures, deal terms or trade secrets.
- Anything covered by professional secrecy, unless you have the client's or patient's consent or another legal permission, and a tool contractually bound to process it only for you.
What this means for confidential data
Practical tip: if you must work with real details, replace them with placeholders ("Patient A", "Client X") before you paste, and check the tool's training and retention settings.
Placeholders help, but they are not the full answer. If a provider may reuse your prompts, it can strip remaining identifiers, and what is left falls outside data protection law. Legally, it can then be pooled and sold. See how on the data trail.
Where Wysor fits
In Wysor you can use the same prompt with GPT, Claude, Gemini and other models in one workspace, and compare answers. We act only as your processor, never train on your prompts or sell them, and every connected model provider runs with zero data retention. You can also save a role and standing instructions in a Persona Agent so you do not have to repeat them.
FAQ
What is a prompt in AI? The text instruction or question you give an AI tool. It tells the AI what to do, with what information, and in what form.
What makes a good AI prompt? Context, a clear task, useful details such as tone and length, and the format you want. Adding limits like "only use the text below" improves reliability.
What is prompt engineering? The practice of writing and refining prompts to get better results. For everyday use, clear instructions and a few rounds of feedback are usually enough.
Can I put patient or client data in a prompt? Only if the tool is approved for it in your organisation and you have a legal basis. Otherwise use placeholders and leave out identifying details.
Further reading
Questions about using AI with sensitive data? Reach us at [email protected] or through the in-app contact form.