Open the chat history of anyone who uses AI at work and you will find the same prompt written ten slightly different ways. A product description on Monday, another one on Wednesday, a third on Friday. Each time the wording drifts a little, and each time the result is a little different. Sometimes better, usually not. If that sounds familiar, our prompt library started for exactly this reason.
The fix is not a better prompt. It is a reusable one: a template that keeps what works and asks only for the parts that change. Below is the process we use for every template in the text prompt collection.
Why one-off prompts don’t scale
When a prompt gives you a great result, it is tempting to think the magic is in the exact words. In practice, the valuable part of a prompt is its structure. Most prompts that work well share the same four building blocks:
- Role: who the model should act as.
- Task: what exactly needs to be produced.
- Constraints: length, tone, things to avoid.
- Output format: a list, a table, plain text, JSON.
Retyped from memory, one of these blocks always goes missing. That’s why the results feel inconsistent, even when the model hasn’t changed at all.
From a good prompt to a template
Start with a prompt that has already produced something you would publish. Here is a typical one:
You are an experienced e-commerce copywriter. Write a product description for our ceramic pour-over coffee set. The audience is home baristas who care about craft. Keep it under 120 words, friendly but not pushy, and end with one short benefit-driven line.
Now mark everything that will change next time. The product, the audience and the word limit are obvious candidates. Replace each one with a named placeholder:
You are an experienced e-commerce copywriter. Write a product
description for {{product_name}}. The audience is {{audience}}.
Keep it under {{word_limit}} words, {{tone}}, and end with one
short benefit-driven line.That’s the whole trick. The structure stays fixed, and the person using the template only answers four questions.
Name your variables like you would name code
Use short, lowercase names with underscores, such as product_name or word_limit. Avoid vague names like text1 or input. Six months from now, a colleague should understand what goes into each field without opening the prompt.
Give every variable a sensible default
A default turns a blank form into a one-click action. If 80% of your descriptions are for the same audience, pre-fill it. People will only change the field when they need to. For fields with a small set of valid answers, like tone, a dropdown beats a free-text field every time.
Decide what stays fixed
Not everything should become a variable. A template with fifteen fields is just a form nobody wants to fill in. Keep the role, the quality bar and the output format fixed, because those are the decisions you have already tested.
A good template makes the easy decisions for you and leaves only the interesting ones.
From our weekly prompt review notes
Test it before you share it
A template is only reusable if it works with values you didn’t think of. Before adding it to your shared library, run a quick check:
- Fill it in with three very different inputs: a cheap product, a premium one and something unusual.
- Read the outputs side by side. Inconsistent length or tone usually means a missing constraint.
- Note which model you tested it with and when. Models change, and so do results.
- Save one good output as a sample, so others know what “good” looks like.
Where to go next
Pick the one prompt you retype most often this week and turn it into a template. Then browse our prompt library for ready-made examples with variables already set up. If you want to dig deeper into prompt structure, Anthropic’s prompt engineering overview is a solid, practical read.