Chatbot training phrases written natively in several languages
Writes the example phrases a support chatbot learns from in each target language, the way real people type them, including mixed scripts, short forms and spelling slips, instead of translating the English phrases word for word. Use it when a bot understands English customers well but misses the same request in Hindi, Tamil, Spanish or Arabic.
The output this prompt produced, unedited.
The prompt
What to give
Intents and what each means: [e.g. track_order: wants to know where the parcel is; cancel_order; change_address; refund_status]
English example phrases per intent: [paste]
Target languages and how customers write them: [e.g. Hindi in Devanagari and in Roman letters, Tamil, Hinglish mixing]
It is something a real customer would type, not a polished translation.
The set mixes short and long forms, questions and statements, and polite and blunt versions.
Mixed-language and romanised forms are included where customers really use them.
Common spelling slips appear in a few phrases, not every one.
No phrase could just as easily belong to another intent; note any that overlap.
Product terms stay as customers say them, even when that is the English word.
Output format
For each intent and language: twelve training phrases.
Overlap warnings: phrases that sit between two intents and how to tell them apart.
Test set: five unseen phrases per intent and language, kept separate from training.
Bot reply per intent in each language, natural and under 40 words, with placeholders such as [order number] untouched.
Notes for a native reviewer: phrases to check for regional usage.
Guardrails
Do not include personal data, real names or real order numbers in phrases.
Keep placeholders and intent names exactly as given, in English.
chatbot training phrasesmultilingual chatbotintent training datachatbot localisationutterance writingmultilingual nlu
How to run this
Copy the prompt and paste it into Gemini, then set the aspect ratio to 3:4. No reference photo is needed for this one.
The prompt is written in Markdown with each part under its own heading, which is what the copy action hands over. Change one section at a time, keeping the rest intact, so you can tell which change did the work.