Prompt Templates: Examples and How to Build One

A prompt template is a prompt with blanks. The fixed text carries everything you have learned about getting a good answer for one kind of task: the role, the format, the limits. The blanks carry what changes each time: the topic, the language, the text to work on. Filling in three blanks is faster than writing a prompt, and it gives you answers of a consistent quality and shape.

The template below explains any programming concept at your level. Type your own values into the fields, switch off a part to see what it contributes, then copy it or open it in a chat app.

Explain a concept at my level

EditCopy

Fill intopiclanguageknownwords

PartsRoleTaskContextFormatConstraints

Role

You are a patient tutor for self-taught programmers.

Task

Explain closures in JavaScript.

Context

I already know variables, functions and arrays. I learn best from a small example I can run myself.

Format

Answer in three parts with bold labels: the idea in one sentence, a short code example with a comment on the key line, and one common mistake.

Constraints

Keep it under 150 words. If you need a term I have not listed as known, explain it in a few words.

Try it

Example replyReplies vary between models and runs.

The idea: A closure is a function that remembers the variables from the place where it was created, even after the outer function has finished running.

Example:

function makeCounter() {
  let count = 0;
  return function () {
    count += 1; // still sees count from makeCounter
    return count;
  };
}

const next = makeCounter();
console.log(next()); // 1
console.log(next()); // 2

Common mistake: expecting two counters to share one total. Every call to makeCounter() creates a new count, so each counter keeps its own.

What makes a good template

A template is only as good as the prompt it came from, so the fixed text should be a prompt you have already run and liked. Three choices separate a useful template from a form that produces mediocre answers faster.

Blank only what changes. If you always want the answer under 150 words, write 150 into the fixed text. Every blank is a decision you have to make again on every use, and a template with ten blanks is slower to fill than writing the prompt fresh.

Give every blank a clear name and a sensible default. {{known}} tells you what to type; {{x}} does not. A default shows the kind of value expected and lets you run the template once without typing anything.

Keep one template per job. A template that explains concepts, reviews code and writes quiz questions depending on a blank does all three worse than three separate templates. The more specific the fixed text, the more it can say about format and standard.

The six parts from how to write a prompt are a good checklist for the fixed text: role, task, context, input, format, constraints. The blanks usually fall in the task, the context and the input.

Prompt template examples

These three cover common everyday jobs. Each has named fields to fill in. For a larger collection of ready prompts, see prompt examples.

Summarize for a specific reader

EditCopy

Fill inaudiencebulletstext

Summarize the text inside the <text> tags for a manager who has two minutes. Use at most 3 bullet points. Lead with any decision or action the reader must take, and leave out background they do not need. <text> Paste the text here </text>

Try it

Write a small function

EditCopy

Fill inlanguagejobempty

Write a Python function that takes a list of numbers and returns the median. Use type hints if the language has them. If the input is empty, raise a ValueError. After the function, add three example calls with their expected results as comments. Return only the code.

Try it

Example replyReplies vary between models and runs.

def median(numbers: list[float]) -> float:
    if not numbers:
        raise ValueError("median() needs at least one number")
    ordered = sorted(numbers)
    middle = len(ordered) // 2
    if len(ordered) % 2 == 1:
        return ordered[middle]
    return (ordered[middle - 1] + ordered[middle]) / 2


# median([3, 1, 2])     returns 2
# median([4, 1, 3, 2])  returns 2.5
# median([])            raises ValueError

Reply to a message

EditCopy

Fill ingoaltonewordsmessage

Write a reply to the message inside the <message> tags. Goal: politely decline and suggest a date next month instead. Tone: friendly and brief. Keep it under 80 words and do not add facts that are not in the message or in the goal. <message> Hi! Would you be free to give a short talk about your project at our meetup on the 14th? </message>

Try it

Example replyReplies vary between models and runs.

Hi, thank you for the invitation, I'm glad you thought of me! Unfortunately I can't make it on the 14th. Would a date next month work instead? I'd be happy to talk about the project then.

Best, [Your name]

The tags around the pasted text are there for a reason. Filled-in text can contain sentences that look like instructions, and delimiters keep the model from confusing your template's instructions with the material it should work on.

How to build your own template

1. Start from a prompt that worked. Take a conversation where the answer came out right, and collect the final version of the prompt, including any corrections you had to add in follow-up messages. Those corrections are the most valuable part, because they fix what the first version got wrong.
2. Mark what would change next time. Read the prompt and underline every word you would edit for the next use. Usually it is two to four things: a subject, an audience, a length, a pasted input.
3. Replace each one with a named blank. Write the blank the way your tool expects it, such as [audience] for a prompt you paste by hand or {audience} for code.
4. Test with three different fillings. Use one typical case, one very short or very long input, and one unusual case. If one of them fails, the fixed text is still assuming something about the first use. Change one sentence at a time, as described in iterating on prompts.
5. Save it where you will find it. A notes file, a text expander or a snippet manager all work. A template you have to search for is a template you will stop using.

Templates in code

In a program, a template is a string with named fields. Python's str.format() fills them in:

SUMMARY_TEMPLATE = """Summarize the text inside the <text> tags for {audience}.
Use at most {bullets} bullet points and lead with any action the reader must take.

<text>
{text}
</text>"""

with open("notes.txt", encoding="utf-8") as f:
    meeting_notes = f.read()

prompt = SUMMARY_TEMPLATE.format(
    audience="a manager who has two minutes",
    bullets=3,
    text=meeting_notes,
)

With str.format(), every literal curly brace in the template has to be doubled ({{ and }}), which matters when the template includes a JSON example. Braces inside the values you pass in, such as code in meeting_notes, need no escaping. Some libraries for building AI apps have their own template classes, but they do the same job as a format string with extra features around it.

When the values come from users, remember that the template sends their text to the model. Keep it inside delimiters, and do not give the model tools or data that a malicious input could misuse.

Template or system prompt

A template and a system prompt solve different problems. The system prompt holds rules that apply to every message in a conversation or an app: the persona, the tone, what the assistant must never do. A template holds one task with blanks. In an app you often use both: a fixed system prompt that sets the ground rules, and a template that builds each user message.