Advanced Prompting — Roles, Examples, Steps, Format

1.Role: decide whose eyes it should see through

If you assign a role, such as "You are an elementary school science teacher," the perspective, tone and level of the explanation change to fit that role. Even for the same topic, photosynthesis, a teacher would explain with analogies and questions, a botanist with chemical equations, and a farmer by connecting it to growing crops. A role is a powerful piece of context that tells the model what kind of text to continue.

But giving a role does not turn the model into an expert in that field. Saying "You are a lawyer" does not create legal knowledge the model doesn't have. Remember that a role changes only the tone and perspective; it does not guarantee factual accuracy.

Roles are also useful for getting criticism. If you say, "You are a strict reviewer seeing this proposal for the first time. Find five weaknesses," you can spot holes you missed before anyone else does.

2.Examples: if it's hard to describe, show it

When the result you want is hard to describe in words, showing one or two examples is the fastest way. Just as people understand at once when they see a sample "along these lines," the model copies the length, tone and structure of the examples. Giving no examples is sometimes called zero-shot, one example one-shot, and several examples few-shot.

Be careful about two things when giving examples. First, the model follows examples very faithfully, so if all three examples are phrased the same way (say, all starting with an "-ing" verb), all the results will be phrased that way too. If you want varied results, give varied examples. Second, so that it doesn't copy the content of the examples as well, say clearly, "Use only the format as a reference; write new content."

3.Steps: break big jobs into pieces

If you hand over a big job like "Write a market research report" all at once, the result tends to be shallow. Breaking it up in the order a person would work makes it much better: first get an outline, then revise the outline yourself, have it write each chapter, and finally have it polish the whole thing. The biggest advantage is that you can check and steer the direction at every step.

For questions that require calculation or reasoning, it also helps to ask, "Don't just give the answer; show the solution step by step." As we saw in Lesson 4, the model writes one token at a time, so having it write out the intermediate process gives the next step something to build on. Above all, it lets a person check each step. Even if the steps look plausible, a wrong calculation can slip in along the way, so check the numbers yourself.

• Step 1: Get an outline or plan first
• Step 2: Revise the plan yourself (remove, add, reorder)
• Step 3: Produce the content part by part
• Step 4: Have it review and polish the whole thing at once
• Step 5: A person checks the facts and numbers

4.Format: fix the shape of the result

If you plan to move the result somewhere else, it is a good idea to define the format precisely. Rather than just saying "as a table," specify the column names and their order too, and you will get results of the same shape even across multiple requests, which makes them easy to compare. Match the format to where it will be used: "in comma-separated format (CSV)" if it is going into a spreadsheet, or "in a title–subheading–body structure" if it is going into a document.

Format also includes length and count. Setting counts, such as "three pros and three cons, one sentence each," keeps it from going on at length about only one side. And if you build rules for marking uncertainty into the format, such as "If you don't know, say you don't know" or "Add (needs checking) to any item you're not sure about," the verification we cover in Lesson 7 becomes easier.

• Wrong: Compare three laptops.Right: Compare laptops A, B and C in a table. Use the columns "Weight, Screen size, Battery life, Price range, Needs checking" in that order. Leave any value you're not sure about blank, and write "needs checking" in the last column.If you define the column names, their order and a rule for marking uncertainty, you can review the result and move it elsewhere as-is.

5.Using the four techniques together

In practice, you mix the techniques. For example, to get a draft of an English email, you might write: "You are a coworker who is good at business English (role). Use the tone of the sample email below as a reference (example). First, show me the key points in three lines, and once I confirm, write the email (steps). Use the order subject, body, closing, and keep the body under 120 words (format)."

When a prompt gets long, don't mix material and instructions; divide it into sections. If you add headers such as "### Instructions," "### Material" and "### Example," or wrap the material in quotation marks, the model won't confuse what is a command and what is content to process. This habit also helps reduce the mistake of following stray instructions that happen to be mixed into the material.

6.When the result isn't what you expected: finding the missing technique

As the tip above says, rather than putting in every technique from the start, it is better to look at the result and add them one at a time. To do that, you need to be able to look at the "symptoms" of a result and guess which technique is missing. The table below lists common symptoms and the matching remedies.

The trick is to try one remedy at a time. If you change several things at once, even if the result improves you won't know what worked, so you are left with no lesson you can reuse the next time you do something similar. Write the remedies that worked into your prompt templates, as in the tip from Lesson 5.