You will be able to decide which background details to include in a request and which to leave out.
Priya works in HR at a logistics company in Tampines. On Monday her manager asks her to tell all staff that outpatient medical claims are moving from paper forms to the company's HR app next month. She opens an assistant and types: write an announcement about the new claims process. What comes back is cheerful, three paragraphs long, full of phrases like "an exciting new chapter", and it could have been written for any company anywhere.
She tries again, this time giving the assistant what she would give a new colleague asked to write it for her. The second draft is close to sendable. The only thing that changed was the background she supplied, which lesson 1.1, The assistant only knows what you put in front of it, called context. This lesson is about which background to give and which to keep back.
If you add only one piece of context, make it the reader. Who are they, what do they already know, and what do they need to decide or do once they have read it? An assistant writes very differently for a finance director with five minutes to spare than for a warehouse supervisor reading on a phone between shifts.
Priya's readers are mostly drivers and warehouse staff. Many read company messages on their phones, and some have used paper claim forms for ten years. They need three things: what changes, from when, and what to do if the app does not work for them.
Once the model knows the reader, it can make many decisions for you. Vocabulary, length and how much to explain all follow from who is reading, so you do not have to spell out every rule. What the reader already knows matters as much as who they are. Tell the assistant if they have seen an earlier version, complained before, or are new to the topic, or it will explain what they know and skip what they do not.
The same facts turn into a different piece of writing depending on where they end up. An email to a client, a slide for your own team meeting, talking points you will read aloud and notes only you will see each need a different shape. The model cannot tell which one you want from the topic alone.
So say it plainly. This is a WhatsApp message for the drivers' group. This is a one-page note my manager will forward to the regional office. These are points I will speak from at tomorrow's briefing. Purpose changes the length, the tone, how much gets explained and what comes first. A group chat message leads with the action, while a note to a regional office usually leads with the reason.
The model has read a great deal about medical claims in general. It has never seen your company's claims policy, does not know what the app is called, and has no way of knowing that the change was announced once last year and then postponed, so staff are doubtful.
Those are the facts that make a draft yours. A good habit is to ask what a new colleague would need to ask you before starting. The usual answers are your product or service, the deadline, the figures involved, what has already been tried and anything that went wrong last time.
Put those in, because when they are missing the model does not stop and ask. It fills the gap with something plausible, and plausible is often wrong in exactly the detail that matters. AI fundamentals lesson 6.1, Why a fluent answer can still be false, explains why.
Here is the context Priya wrote. I work in HR at a logistics company in Singapore. This message is for our drivers and warehouse staff, who mostly read it on their phones and have used paper claim forms for years. From 1 July, outpatient claims must go through the HR app instead of paper. We announced this last year and postponed it, so some staff will doubt it is happening. Anyone who has trouble with the app can come to the HR counter at the depot on Tuesday or Thursday afternoons.
That is five sentences, and every one of them changes the draft.
Good context is selective. The task needs the reader, the purpose and the facts that shape the answer. It does not need client names, NRIC numbers, bank account details, salaries or anyone's medical history, and typing those into a chat puts a copy of them somewhere outside your control.
Priya's paragraph names no employee and no clinic. If she later wants help replying to one driver whose claim was rejected, she can describe the situation as "a driver whose claim was rejected because the receipt photo was unclear" and add his name herself when she sends the reply. Placeholders such as [client name] or [amount] work well for this. The model writes around them and you fill them in at the end.
A simple test helps here: if a detail would not change the draft, it does not belong in the prompt. Your company may also have rules about what can go into which tool, and lesson 8.2, Data settings and the rules at your workplace, shows you how to find them.
For most everyday tasks, three to five sentences of context is enough. A sensitive message or a long report might need more. If you find yourself writing two pages of background, the job may be too big for one request, which is what module 3 is about.
The real saving comes from tasks you repeat. Write the context once for a job you do every week and you can reuse it, changing only the details each time. Think of the task you return to most often, then ask the three questions from this lesson about it: who reads it, where it ends up, and what you know that the model cannot.
Write a context paragraph of three to five sentences for a task you do often, covering reader, purpose and the facts only you know.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).