Prompt Templates: Build Reusable AI Prompts

A prompt template is a reusable prompt with placeholders that get swapped out for different inputs. Instead of writing a fresh prompt every time, you design one template that works across dozens — or thousands — of use cases.

From One-Off to Reusable

Real-world usage rarely means crafting one prompt at a time. You might need to summarize 500 customer tickets, extract data from 200 resumes, or classify emails all day. Writing a unique prompt for each is impractical.

Templates separate the instructions (what to do) from the input (what to do it on):

Summarize the following support ticket in 2-3 sentences.
Focus on the customer's core issue and urgency level.

Ticket: {{TICKET_TEXT}}

The {{TICKET_TEXT}} placeholder is a variable — it gets replaced with actual content at runtime. The instructions stay constant; only the input changes.

Anatomy of a Good Template

Effective templates have three layers:

1. Role and context — Who the model is and what it’s doing
2. Instructions — The task, constraints, and output format
3. Input variables — Clearly marked placeholders for dynamic content

<role>You are a senior technical recruiter.</role>
<instructions>
Evaluate the candidate profile below. Rate fit on a scale
of 1-5 for each requirement. Return results as JSON.
</instructions>
<requirements>{{JOB_REQUIREMENTS}}</requirements>
<candidate>{{CANDIDATE_PROFILE}}</candidate>

Using delimiters to separate these layers makes templates easier to read, debug, and maintain.

Including Examples in Templates

Templates become more powerful with embedded few-shot examples — sample input-output pairs showing the model exactly what you expect:

Classify the support ticket into one of: billing, technical, account, other.

<examples>
<example>
Input: "I was charged twice for my subscription"
Output: billing
</example>
<example>
Input: "The app crashes when I open settings"
Output: technical
</example>
</examples>

Ticket: {{TICKET_TEXT}}

Two to five diverse examples typically give the best balance of accuracy and token efficiency.

Tips

• Name variables descriptively — {{CUSTOMER_FEEDBACK}} is clearer than {{INPUT}}
• Version your templates — small wording changes can shift output quality, so track what changed and when
• Test with edge cases — try your template with unusual inputs to catch failure modes early

Templates handle single tasks well. But what about tasks too complex for a single prompt? That’s where chaining comes in.