Give it something to work with.
How to hand a model the material it needs: what to paste, where to put it, how to fence it off, when to show an example, and what to leave out.
Someone asks a colleague, "Can you reply to this?" and then does not hand over the email. The colleague would ask for it. A model will not. It will write a reply to an email it has never seen, and the reply will be fluent, polite and useless.
The last lesson named the parts of a prompt. This one is about the biggest part by volume, the material, and the small mechanics of handing it over so the model can actually use it.
The model only knows what you show it
Everything the model can use sits in the context window, the stretch of text it can see at once: your messages, its replies, and anything you pasted or attached. It has no access to your inbox, your files or your last meeting unless the app puts them there. So the single biggest upgrade to most prompts is pasting the thing. The email thread, not a description of it. The draft, not "my draft is a bit formal." The actual figures, not "the numbers from last quarter."
This works because models are far more reliable when reading than when remembering. Asked to summarize a document in front of it, a model stays close to the text. Asked about the same document from memory, it reconstructs, and reconstructions drift. Giving it the text moves the task from the weak side of the profile to the strong side.
Long material first, question last
Order matters more than it seems. When you paste something long, put the material first and your instruction at the end. The instruction nearest the end of the prompt tends to have the strongest pull on what the model does next. If you lead with a question and follow it with three pages of pasted text, the question can get buried. Lead with the pages and finish with Based on the thread above, draft a reply that..., and the model arrives at the task with everything already in view.
For short material the order barely matters. For anything over a few paragraphs, make it a habit.
Keep instructions and material apart
When you paste text into a prompt, the model has to work out which words are instructions for it and which are material to work on. Usually it manages. Sometimes it does not: an email thread that itself contains "please summarize this," or a document with its own headings, can blur the line. The fix is a visible boundary.
Simple delimiters work. A delimiter is any marker that shows where one thing ends and another begins: a line of three dashes, a pair of labels like <email> and </email> wrapped around the pasted text, or just EMAIL STARTS and EMAIL ENDS. Then refer to the material by name: Using the email above, list every question the customer asked. Angle-bracket tags look technical, but they are only labels. Models have seen millions of them and treat them as clean boundaries. When you paste more than one thing, label each: <draft>, <style_guide>, <data>.
One example beats a paragraph of description
People try to describe the output they want: professional but approachable, concise, with a clear call to action, not too salesy. Each of those words means something different to every reader, and to the model. A single example does the same job with no ambiguity. Paste one email you were happy with and say Write the new one in the same style as this. Want a specific layout for meeting notes? Paste last week's notes. Want a product description shaped a certain way? Paste one that works.
This is sometimes called few-shot prompting, where a shot is just an example: one-shot is one example, few-shot is a handful. One good example usually gets you most of the benefit. Two or three help when the pattern is subtle. Pick examples that are genuinely like what you want, because the model copies the example you gave, not the one you had in mind.
Roles, and saying what to do
A line like You are an experienced tax accountant or Act as a pediatric nurse explaining this to a worried parent does one useful thing: it shifts the model toward the vocabulary, concerns and level of detail of that kind of person. A role will not make the model know more than it knows, and it will not make an answer correct. But it changes which patterns the model reaches for, and when a field has its own language, that matters.
Alongside the role, phrase instructions as what to do rather than what not to do. Do not be formal leaves the model choosing among a hundred alternatives. Write like a friendly colleague in a quick chat message picks one. Negatives also keep the unwanted idea in view, which is not where you want it. Say the target, not the thing to miss.
What to leave out
More material is not always better. Three things do not belong in a prompt.
- Secrets. Passwords, keys, account numbers, anything you would not paste into a form on a website you do not control. What you type travels to the provider's servers. Replace them with placeholders like
[API_KEY]. - Other people's private data. A colleague's medical note, a customer's home address, a candidate's salary history. If the task does not need the identifying details, remove or blur them before pasting. (A later lesson covers privacy properly.)
- Anything irrelevant. The forty-message thread when only the last three matter. The whole spreadsheet when the question is about one column. Irrelevant text is not neutral: it is more for the model to read, and the detail you care about can get lost in it. Trim to what the task needs.
The test is simple. If a sharp colleague would need it to do the task, paste it. If they would not, leave it out.
Try it
Find an email you need to answer, or any longer message you have. In a chatbot, paste it between <email> and </email>, then below it type: Using the email above, list every question or request the sender made, then draft a reply that answers each one in two or three sentences, in the tone of a friendly colleague. Compare that with what you get from Help me reply to an email about the project delay with nothing pasted. Then try once more with a past reply of yours pasted as an example and match the style of my earlier reply added at the end.
Key terms
- Delimiter
- A visible marker, such as a line of dashes or a pair of tags, that separates your instructions from pasted material.
- Few-shot prompting
- Giving the model one or more examples of the output you want so it can copy the pattern.
- Role
- A line that tells the model what kind of person to answer as, which shifts its vocabulary and focus.
- Placeholder
- A stand-in such as [NAME] or [API_KEY] used in place of a real secret or private detail.
Check yourself
Paste the actual material, put it first with the question last, fence it off with a label, and show one example of good rather than describing it.