Build a small set of examples you can reuse to keep a repeated task consistent.
For a task you do every week, the most reliable way to get consistent output is a saved prompt built around three real examples of your own work. You label the examples clearly, test the prompt on new inputs and adjust the examples until the results come out the same way each time.
Some tasks come round every week but look slightly different each time, such as replies to the same five customer questions, notes from the Monday team meeting or a short Friday update to your manager. You have probably written dozens of these and know what a good one looks like, yet when you start from a blank prompt each time, the result comes out slightly different from one week to the next.
Daniel runs customer service for an online furniture shop in Singapore (an invented example). About a third of Daniel's inbox is "where is my order" messages, and the shop's part-timers answer them in three different styles.
Pick a task you do at least weekly and where consistency matters. Good candidates are replies to common queries, meeting notes in a set format, short status updates, product descriptions and job ad drafts. A task you do once a quarter is a poor choice, because you won't reuse the prompt often enough to repay the effort. The task also needs a clear input and a clear output. For Daniel, the input is a customer message asking where an order is, and the output is Daniel's reply to that customer.
Next, collect three good examples. Each one must be a real output you were happy with, paired with the input it answered. Invented examples tend to be cleaner and more uniform than real work, and the model copies that tidiness into its own answers. What you learned in Lesson 2.1, "One good example beats a paragraph of description", about choosing examples applies here.
The three examples should differ from each other in useful ways, such as mood and situation, while sharing the same voice and structure. Daniel picks three replies, which cover different moods and situations in one voice and one structure:
a reply to a polite customer whose order was on schedule a reply to an annoyed customer whose sofa was a week late a reply to a customer who had given the wrong unit number
Before you use your examples, remove anything confidential. Replace names with [customer name], order numbers with [order number] and addresses with [address]. Then check for other identifying details, such as a phone number in a signature. The examples only need to show the pattern, so placeholders do the job as well as anyone's real details would.
A reusable example prompt is a saved prompt with four parts, in this order. First comes a short paragraph of context and the task. Second come the examples, each clearly labelled, with input and output visibly separate. Third come any rules about what to avoid. Fourth is a clear marker that the examples have ended and the new input starts.
Base the opening paragraph on the model from Module 1. Daniel's prompt opens with two sentences. The first says that Daniel runs customer service for an online furniture shop in Singapore, and the second asks the model to draft a reply to the customer message at the end, following the pattern of the three examples.
Labels are markers such as "Example 1. Customer message:", "Reply:" and "New customer message:" that separate inputs, outputs and the new input. Without them, a model may mistake an example for part of the new message, or treat the new message as a fourth example and answer all four. Daniel writes "Example 1. Customer message:" followed by the message, then "Reply:" followed by the reply, and uses the same layout for Examples 2 and 3.
After the examples, Daniel adds the habits to avoid from Lesson 2.2, "Style samples and examples of what to avoid", which is also the source of what you apply here about voice. The prompt ends with "New customer message:" and a blank where the new input goes.
Test the prompt on three new inputs that you didn't use as examples, and use a fresh chat for each one. Read the outputs as a set and ask two questions: do they follow the pattern, and is anything copied that shouldn't be there?
The most common problem is over-copying, where the model reproduces details, phrasing, openings, product names or length from an example in places where they don't belong. You may see the same opening line every time, the same product name turning up in unrelated replies, every output coming out the length of the longest example, or details from one example appearing where they don't apply. It usually happens because one example dominates the others, for instance because it is the longest and most detailed of the three.
In Daniel's first round of tests, every output mentioned "a one-week delay", even for an order that was on time. The late-sofa example was the longest and most detailed, and it dominated the other two. Daniel replaced it with a shorter late-delivery reply, and in the next round the outputs matched each customer's situation.
Fix over-copying by replacing or trimming the example that is causing it. Adding an instruction to fight it rarely works, because the example is usually stronger than the instruction. After each round, write down what you changed and why. Two rounds are usually enough. If a third round still shows problems, check again whether your three examples really share the pattern you want the model to follow.
At the end you have one saved prompt and a short test record. The prompt holds the context, three labelled examples, the rules about what to avoid and a marked slot for the new input. The record lists the three test inputs, one line on how each output did, and what you changed. Daniel's record is six lines long. The part-timers now paste the same prompt, and the replies sound as though the same shop wrote them.
Keep the prompt somewhere easy to find, because Module 9 turns prompts like this one into templates you can reuse across tasks.
For now, decide which weekly task you will use and go and find your three examples.
Save a prompt with three labelled examples, test it on three new inputs, and write down what you changed after testing.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).