Persona & Context: Role-Playing and The Art of Context Management

Welcome to Module 4. We've covered structure and reasoning. Now, we dive into Persona & Context. This module is about who the model is pretending to be and what information it has access to.

Assigning a persona to an LLM changes its default behavior significantly. It shifts the probability distribution of tokens towards a specific domain, tone, or expertise level.

• Tone: "Explain like I'm 5" vs "Explain like a PhD Physics Professor".
• Expertise: "Act as a Senior React Developer" vs "Act as a Junior Python Developer".
• Style: "Write in the style of Shakespeare" vs "Write in the style of a technical manual".

Prompt:

You are a world-class copywriter for a luxury brand. Write a product description for a simple white t-shirt. Output: "Elevate your everyday with the purity of organic cotton. Meticulously crafted for an effortless silhouette..."

Prompt:

You are a chaotic goblin. Describe a white t-shirt. Output: "Shiny white cloth! Soft! Good for hiding crumbs! Want!"

Often, models hallucinate or bring in outside knowledge when they shouldn't. The best way to combat this is to limit their scope within the persona.

Prompt:

You are a customer support agent for Acme Corp. Answer ONLY based on the provided FAQ. If the answer is not in the FAQ, say "I don't know". Do not use outside knowledge.

When working with large documents or retrieved information (RAG - Retrieval Augmented Generation), context management becomes critical.

LLMs are great at remembering the beginning and the end of a long prompt but tend to "forget" details in the middle.

Strategy:

• Put Key Instructions at the Start: Tell the model what to do with the context before giving it the context.
• Put the Question/Task at the End: Remind the model of the specific question after the context block.

Bad Prompt Structure:

[Huge context dump...] Summarize this.

Good Prompt Structure:

You are a summarization assistant. Your task is to extract key dates from the text below.

Text: """ [Huge context dump...] """

Task: Extract all dates from the text above.

• Stuffing: Pasting the entire document into the prompt.
• RAG: Using a database to find only the relevant chunks of text and pasting those into the prompt.

For massive contexts (books, codebases), RAG is essential. But for shorter contexts (articles, emails), stuffing is often better because the model sees the full picture.

In the next module, we will explore Evaluation & Optimization, learning how to measure if our prompts are actually working.