AI prompts: a guide to writing them, plus 50 ready-to-use work prompts

Good AI prompts read like a clear brief to a capable new colleague. Say what you want produced, who it is for, why it matters, what material to use and the format you need back, then check the answer and refine it. This guide gives the method, a worked example and 50 work prompts to copy.

The method draws on the prompting guides that OpenAI, Google, Microsoft and Anthropic publish for their own assistants, all read on 14 October 2026. The prompts work in ChatGPT, Gemini, Copilot, Claude or any similar assistant, in Arabic or English. O AI is not affiliated with OpenAI or any other company named here.

What is a prompt, and why do AI prompts change the answer?

OpenAI describes a prompt as the input that starts a conversation or triggers a response from a language model. It is usually text, but it can also be an image or audio [1]. The model works only from that prompt and whatever you attach. It does not know your company, your customer or the tone your manager expects unless you tell it.

OpenAI's help centre suggests writing to the model as you would to a person: "do this task for me" without context rarely gets what you wanted, while clear, detailed instructions raise the chance of success [2]. Google adds: write as if speaking to another person, in short, specific sentences [3].

Anthropic puts it most memorably. Treat the assistant as a brilliant but new employee who lacks context on your norms and workflows, and apply a "golden rule": if a colleague with no background would be confused by your prompt, the model will be too [5].

The five parts of a good work prompt

The four vendors describe the same building blocks under different names. Use whichever vocabulary your team already knows.

Not every prompt needs all five: Google says its four areas are not all needed every time, and Microsoft says only a clear goal is required [3][4]. For a quick internal task, task and source are often enough; for anything that leaves the company, add context and format.

Context works best when it includes the reason. Anthropic's example: instead of a bare "never use ellipses", explain that the answer will be read aloud by a text-to-speech engine that cannot pronounce them, and the model generalises from the explanation [5]. In an office, "keep it under 80 words because the client reads email on his phone" works better than "keep it short".

Before and after: one prompt rewritten

Illustrative example. A building-materials supplier in Dammam receives an angry email from a contractor whose order arrived late.

Weak prompt: "Write an email to a customer about a late delivery."

The answer will be generic: invented reasons, an unclear tone, perhaps a promise the company cannot keep.

Strong prompt: "You are the customer service lead at a building-materials supplier in Dammam. Draft a reply to the complaint below from a contractor whose order of [product] arrived [number of days] days late. Apologise once, explain the cause in one sentence ([cause]), offer [remedy] and give a direct contact ([name and phone]). Formal but warm, in the language of the complaint, under 120 words. Do not promise any date I have not given you. Complaint: [paste]."

Each part removes a guess. The role sets vocabulary and formality, the single verb stops the model drifting into analysis, the context says what matters, the pasted complaint keeps names accurate, and the format makes the reply usable without heavy editing.

Seven techniques that make AI prompts more reliable

1. Give examples of what good looks like. Anthropic calls examples one of the most reliable ways to steer format, tone and structure, and suggests 3–5 that are relevant and varied [5]. Paste three past letters your manager approved and ask for "the same style".
2. Put long documents first and the question last. For long inputs, Anthropic recommends placing documents at the top and the query at the end; in its tests this improved response quality by up to 30% on complex, multi-document inputs [5].
3. Ask for quotes before conclusions. For long files, ask the assistant to quote the relevant passages first and then answer [5]. You can check each quote against the page.
4. Separate instructions from material. Label each part, for example "Instructions:", "Document:", "Glossary:". Anthropic recommends tags for the same reason: they stop the model confusing your instructions with the text it is working on [5].
5. Break big jobs into steps. OpenAI advises splitting a complex task into smaller, focused prompts and asking for several options when you are unsure [2]. Summarise the tender first, then list the risks, then draft the questions to the client.
6. Name the tone. OpenAI suggests descriptive adjectives such as formal, friendly or professional [1].
7. Refine in the same conversation. OpenAI treats the first prompt as a draft to adjust after you see the answer [1], and Microsoft notes that the same prompt can produce different responses, so expect some back-and-forth [4].

One more habit, from practice rather than a vendor guide: end important prompts with "Before you start, ask me up to three questions if anything is missing." Gaps surface before the model fills them with guesses.

50 ready-to-use AI prompts for work, by task

Each prompt is written for a company in Saudi Arabia. Replace everything in [brackets], paste your material where it says [paste] or attach the file, and check the result with the checklist further down. Every prompt also works in Arabic: add "Reply in formal Arabic" or "Reply in Saudi dialect" at the end.

Email and correspondence

Meetings, reports and documents

Sales and customer service

Marketing and social media

HR and administration

Finance, spreadsheets and data

Research, translation and planning

When a prompt from this list runs every day on the same kind of input, it has become a workflow. Our guide to AI agents, RPA and workflow tools explains when it is worth automating that step.

What should never go into a prompt?

The commonest privacy mistake is pasting a whole customer file "for context". Saudi Arabia's Personal Data Protection Law defines personal data broadly: any data that identifies a person directly or indirectly, including a name, ID number, address, contact numbers, bank account and card numbers, and photos [8]. The same law restricts transferring personal data outside the Kingdom and limits any transfer to the minimum personal data needed [8].

If an assistant processes your data outside Saudi Arabia, pasting a customer list into it can amount to such a transfer. Our guide to data residency in Saudi Arabia explains the conditions. This section is general information; for a specific case, check SDAIA's guidance or ask a licensed advisor.

The plan matters too. On ChatGPT's Free, Plus and Pro plans, OpenAI may use your content to train its models unless you turn off "Improve the model for everyone" under Settings > Data controls; content from ChatGPT Business, Enterprise and Edu is not used for training by default [9].

Three working rules:

• Replace identifiers with labels: "Customer A", "Employee 2", "[ID removed]". The model rarely needs the real name to draft a reply.
• Paste the minimum: the clause in question, not the whole contract; the three relevant rows, not the full payroll.
• Use the company's business account for internal material, never a personal free account.

Check every answer before you use it

A fluent answer is not a correct one. OpenAI warns that its models can produce incorrect dates and facts, fabricated quotes and citations to sources that do not exist, and can sound confident while wrong; it recommends treating answers as a first draft and verifying important information [7]. Microsoft gives the same advice for Copilot [4].

Run this check before anything leaves your desk:

1. Figures and dates: every number matches the source you gave, and totals are recalculated.
2. Names and titles: spelled exactly as in your records, with the right honorific.
3. References: every regulation, article number or link was opened by a person, preferably on a .gov.sa page.
4. Promises: no commitment on price, date or scope that you did not authorise.
5. Tone: read it aloud; it should sound like your company, not like a template.
6. Language: Arabic reads natively, with no literal translation of English idioms.

If the same error keeps coming back, fix the prompt rather than the output: add the missing context, an example or a rule.

Save your best prompts and reuse them

A prompt that works is a company asset. Two ways to keep it:

• Custom instructions. In ChatGPT, custom instructions apply to every chat and suit standing preferences such as "write in formal Saudi business Arabic" or "use SAR and the Gregorian date first". Free and Go users can save up to 1,500 characters; Plus, Pro and business plans up to 5,000 [10]. OpenAI notes that information from custom instructions is also used to improve its models, unless you have opted out [10].
• A shared prompt library. A simple spreadsheet with five columns: task, prompt, owner, date last tested, and notes on what to check. Anthropic recommends deciding what a good result looks like, and how you will test it, before refining a prompt [6]; record that in the notes column.

Review the library every quarter: assistants change their models often, and a prompt that worked in January may need a new example by June.

Your next step

This week, pick three AI prompts from the list above that match tasks your team already does. Fill in the brackets, run each on a real but anonymised sample, and keep the version that needed the fewest edits in a shared library. Our guide on how AI can help your business shows how to choose one task and measure the time saved before you spend on tools. You can also read about O AI, a software and AI company in Al Khobar, on our home page.