Text classification prompt template

Paste this into a chat window and replace each {{variable}}, or open it in the prompt builder to fill them in a form. Sections are marked with XML tags, which every major model reads well.

You classify {{item_type}} into exactly one category. You are consistent: the same text always gets the same label.

<examples>
<example>
<input>
I was charged twice for March and only have one account.
</input>
<output>
{"label": "billing", "reason": "Asks about a duplicate charge."}
</output>
</example>
<example>
<input>
The export button does nothing when I click it in Firefox.
</input>
<output>
{"label": "bug", "reason": "Reports a feature that doesn’t work."}
</output>
</example>
<example>
<input>
Hi
</input>
<output>
{"label": "unclear", "reason": "Too short to tell what the customer needs."}
</output>
</example>
</examples>

<constraints>
Categories:
{{labels}}
- unclear: the text doesn’t fit any category, fits several equally, or is too short to judge.

Rules:
- Choose by the main intent of the text, not by keywords that appear in it.
- Use “unclear” rather than forcing a weak match.
</constraints>

<output_format>
JSON only: {"label": "<one category>", "reason": "<one short sentence>"}
</output_format>

<task>
Classify this {{item_type}}:

<text>
{{text}}
</text>
</task>

You classify {{item_type}} into exactly one category. You are consistent: the same text always gets the same label.

<constraints>
Categories:
{{labels}}
- unclear: the text doesn’t fit any category, fits several equally, or is too short to judge.

Rules:
- Choose by the main intent of the text, not by keywords that appear in it.
- Use “unclear” rather than forcing a weak match.
</constraints>

<output_format>
JSON only: {"label": "<one category>", "reason": "<one short sentence>"}
</output_format>

<examples>
<example>
<input>
I was charged twice for March and only have one account.
</input>
<output>
{"label": "billing", "reason": "Asks about a duplicate charge."}
</output>
</example>
<example>
<input>
The export button does nothing when I click it in Firefox.
</input>
<output>
{"label": "bug", "reason": "Reports a feature that doesn’t work."}
</output>
</example>
<example>
<input>
Hi
</input>
<output>
{"label": "unclear", "reason": "Too short to tell what the customer needs."}
</output>
</example>
</examples>

<task>
Classify this {{item_type}}:

<text>
{{text}}
</text>
</task>

{{item_type}}

What is being classified, such as “support ticket”, “app review” or “commit message”.

{{text}}

The text to classify.

{{labels}}

One line per category with a definition, such as “- billing: charges, invoices, refunds, plan changes”. Definitions matter more than names.

{{item_type}}

support ticket

{{text}}

Since yesterday I can’t sign in with Google, it just loops back to the login page. Other people on my team are fine.

{{labels}}

- billing: charges, invoices, refunds, plan or seat changes
- bug: something in the product doesn’t work as documented
- how-to: the customer wants to know how to do something that works
- account: login, password, SSO, deleting or merging accounts

What to look for. This one sits between “bug” and “account”. Under “main intent” it is most likely account (a login problem for one user), and the reason sentence should say so. If your team would route it to engineering instead, sharpen the definitions; that’s the policy decision the template makes visible.

You classify support ticket into exactly one category. You are consistent: the same text always gets the same label.

<examples>
<example>
<input>
I was charged twice for March and only have one account.
</input>
<output>
{"label": "billing", "reason": "Asks about a duplicate charge."}
</output>
</example>
<example>
<input>
The export button does nothing when I click it in Firefox.
</input>
<output>
{"label": "bug", "reason": "Reports a feature that doesn’t work."}
</output>
</example>
<example>
<input>
Hi
</input>
<output>
{"label": "unclear", "reason": "Too short to tell what the customer needs."}
</output>
</example>
</examples>

<constraints>
Categories:
- billing: charges, invoices, refunds, plan or seat changes
- bug: something in the product doesn’t work as documented
- how-to: the customer wants to know how to do something that works
- account: login, password, SSO, deleting or merging accounts
- unclear: the text doesn’t fit any category, fits several equally, or is too short to judge.

Rules:
- Choose by the main intent of the text, not by keywords that appear in it.
- Use “unclear” rather than forcing a weak match.
</constraints>

<output_format>
JSON only: {"label": "<one category>", "reason": "<one short sentence>"}
</output_format>

<task>
Classify this support ticket:

<text>
Since yesterday I can’t sign in with Google, it just loops back to the login page. Other people on my team are fine.
</text>
</task>

• All models

  Set the model’s output to a schema with an enum of your labels using the provider’s structured outputs. The model can then only return a label you defined.

• Claude

  Anthropic notes that structured outputs don’t guarantee the capitalisation of enum values. Compare labels case-insensitively in your code, or keep them all lower-case as here.

• OpenAI GPT

  For high-volume classification, a small, fast model is often enough once the examples are good. Price your volume with the cost calculator before choosing.

• Extract structured JSONPull fields from emails, documents or pages into a fixed JSON shape, with null instead of guesses.
• LLM-as-judge with a rubricGrade one answer against a written rubric and a reference answer, pass or fail per criterion, with evidence.
• Generate eval test casesDraft realistic, edge-case and adversarial inputs for your prompt, each with the expected behaviour.

Building this into an app that sends the same instructions with every request? Start from a system prompt instead; each comes with design notes, failure modes and a monthly cost. In the system prompt library: Support ticket triage and routing or Content moderation classifier.

Edited the template? Compare your version with this one in the prompt diff, and count it for your model with the token counter. In the glossary: few-shot prompting, structured outputs. Browse all 29 prompt templates.