Making a Prompt into a Template
I understand the approach of separating changing parts of a prompt as variables and fixing the standards that should not change as a template.
Prompt Templates: From One-Off Sentences to Reusable Structures
When you first learn prompting, you focus on writing a single sentence well.
But once you start repeating the same task multiple times, the problem changes. A good answer today comes out differently tomorrow, and as people request it with slightly different wording, the results fluctuate.
The problem isn't that you can't write prompts. It's that you're writing a new one each time.
Repeating words fluctuate each time
Prompts written that way are bound to that person, that moment, that intuition. Write it again next time and it changes slightly; have someone else write it and it changes more. The reason results fluctuate is not because of the model, but because the input changes each time.
If there's an answer that came out well once, there's a reason behind that prompt. The role was well-defined, or the context was sufficient, or the output format was clear. A template is what lets you use that reason again next time.
What changes and what shouldn't
A prompt has parts that change and parts that shouldn't.
Topic, target audience, output length, input data — these can vary by task. But role, judgment criteria, output format, and constraints should be maintained repeatedly. A template starts from distinguishing these two things.
When you pull out the changing parts as variables, the prompt becomes much easier to handle.
You are a technical writer explaining {domain}.
Topic: {topic}
Target audience: {audience}
Reader's current level: {level}
{topic}, {audience}, {level} change each time. But the role and judgment criteria stay fixed. With just this one distinction, a prompt transforms from a one-time sentence into a repeatable structure.
Templates don't confine thinking
There's a misconception that templates make answers predictable.
It's actually the opposite. A template isn't a device to make AI's answers identical. It's a device to separate parts that can change from parts that shouldn't.
Without a structure, the AI reinterprets everything. The role, criteria, and format — it decides them all from scratch with each input. With a structure, the AI only creates variations within established criteria. Even when the input changes, the result's structure is maintained.
A prompt template isn't for mechanically repeating words. It's a device for fixing criteria that fluctuate each time.
Good structure creates freedom
A good template contains five things.
Role, variables, output format, judgment criteria, and exception handling method.
In practice, it looks like this.
[ROLE]
You are a technical writer in the {domain} field.
[CONTEXT]
Topic: {topic}
Target audience: {audience}
Reader's current level: {level}
[CRITERIA]
A good answer must satisfy the following criteria:
- The reader should be able to act on it immediately.
- It should present specific judgment criteria rather than abstract advice.
- Uncertain content should be marked as speculation.
[CONSTRAINTS]
- Specialized terms should be explained the first time they appear.
- Don't guess about unknown content.
- Write in {paragraph_count} paragraphs or fewer.
[FALLBACK]
If information is insufficient, don't fill in arbitrarily—ask what additional information is needed first.
[OUTPUT]
{output_format}
{topic}, {audience}, {level}, {paragraph_count} change each time. But the role, constraints, judgment criteria, and output method stay fixed. Results aren't completely standardized, but they're created within the same criteria each time.
The more fixed the criteria, the less the answer's direction wavers even when variables change.
Filling in the blanks completes it
The template's advantage becomes clearer when you actually fill in real values.
Just filling in the changing values from the structure above:
{domain} = startup marketing
{topic} = ways to reduce CAC in early-stage startups
{audience} = first-year non-developer founder
{level} = knows marketing basics but is new to data analysis
{paragraph_count} = 4
{output_format} = draft under 800 characters, with 3-line summary at the end
When you put these values into the template, it becomes a finished prompt you can paste directly into Claude, ChatGPT, or Gemini.
[ROLE]
You are a technical writer in the startup marketing field.
[CONTEXT]
Topic: Ways to reduce CAC in early-stage startups
Target audience: First-year non-developer founder
Reader's current level: Knows marketing basics but is new to data analysis
[CRITERIA]
A good answer must satisfy the following criteria:
- The reader should be able to act on it immediately.
- It should present specific judgment criteria rather than abstract advice.
- Uncertain content should be marked as speculation.
[CONSTRAINTS]
- Specialized terms should be explained the first time they appear.
- Don't guess about unknown content.
- Write in 4 paragraphs or fewer.
[FALLBACK]
If information is insufficient, don't fill in arbitrarily—ask what additional information is needed first.
[OUTPUT]
Write as a draft under 800 characters, and summarize only the key points in 3 lines at the end.
The template's structure stays the same. What changed are only the variables.
Instead of writing the prompt from scratch each time, you can replace only the changing values and reuse the same criteria for different tasks.
Repeatable prompts
Making a prompt into a template isn't about repeating the same thing to the AI.
It's about making the criteria you value reusable. The moment you decide which parts to change and which to hold onto, a prompt transforms from a simple sentence into a tool.
Variables reveal what changes. Templates hold onto what shouldn't.