Think Locally, Systematize in English

Many Turkish businesses ask: "If we prompt in English, how do we generate Turkish content for our customers?" The solution is not to prompt in Turkish. The solution is to separate the Logical Instruction Layer from the Generative Output Layer.

Decoupling System and Content

In a professional workflow, the prompt that instructs the model (system prompt, variable checks, JSON format requirements, safety rules) should be written in English. Within that English prompt, you explicitly command the model to output the final user-facing text in localized Turkish.

This allows the model's core reasoning engine to operate at peak efficiency in English, while its translation weights are only used at the final output generation boundary.

Operating Model: Speak culturally to the market. Build structurally in English. Hybrid prompt everything.

A Practical Hybrid Prompt Blueprint

Here is how a hybrid prompt for a financial transaction notification should be structured:

# SYSTEM INSTRUCTION (English)
You are an institutional financial compliance auditor.
Analyze the transaction payload below.
Check if the transfer exceeds limits or violates local AML rules.

# OUTPUT SPECIFICATIONS (English)
- Output a valid JSON payload.
- Key "status": either "APPROVED" or "FLAGGED".
- Key "reason_tr": Write a detailed, professional explanation in Turkish 
  for the final user explaining the decision, matching brand guidelines.

# TRANSACTION PAYLOAD (JSON)
{ "user_id": "usr_99", "amount": 150000, "currency": "TRY" }

By using this layout, the reasoning, structure, and constraints are processed in English, avoiding logic drift and saving token overhead, while the customer receives copy in native, brand-aligned Turkish.