LLM Prompts for News Processing

Overview

The system uses various prompts for different stages: metadata generation, RAG retrieval, synthesis, composition, and chat response.

Metadata Generation Prompts

For each article, extract:

• Summary (2-3 sentences)
• Sentiment (positive/negative/neutral + score)
• Keywords (5-10)
• Topics

Prompt template: "Summarize this article in 2-3 sentences. Analyze sentiment. Extract keywords and topics."

RAG Synthesis Prompts

Given a collection of articles and query: "Given the following article metadata, retrieve and synthesize the most relevant information for creating a news segment on [query]."

Composition Prompts

Using selected persona: "You are [persona_name]. Compose a 300-word news segment based on this synthesized information. Follow these guidelines: [guidance]. Maintain this tone: [tone_words]."

Chat Adjustment Prompts

Process user input to modify composition: "Adjust the current persona parameters based on this user instruction: [user_message]. Update temperature, guidance, or tone as appropriate. Return the modified configuration."

Prompt Engineering Standards

• Use clear instructions
• Include few-shot examples where needed
• Specify output format (JSON/XML)
• Limit token usage for efficiency

---

Checklist

• Prompts designed
• Prompt performance tested

Ledger

Prompt Type	Template	Performance Note
Summary	"Summarize..."	Good accuracy
RAG	"Retrieve..."	Needs tuning
Compose