Prompting Basics

Say what, for whom, in what shape

Treat a prompt like a brief to a capable colleague who knows nothing about your situation. Include: the task, the audience, necessary context, constraints, the exact output format, and one or two examples (few-shot) of good output. A system prompt sets durable role and rules; the user message holds the specific request. Ask the model to say "I don't know" when information is missing, and to quote its sources when you supply documents. Iterate: change one thing at a time and compare results on a fixed set of test inputs.

Prompt, ground, act

Good prompts, retrieved facts and tools make a model useful in a product.

A structured prompt

A template that states role, task, rules, format and an example. Placeholders are in braces.

You are a support assistant for {product}.
Task: classify the ticket below and draft a reply.
Rules:
- Use only the policy text provided; if it does not cover the case, say so.
- Keep the reply under 80 words.
Output JSON: {"category": "...", "reply": "..."}

Example ticket: "I was charged twice."
Example output: {"category": "billing", "reply": "Sorry about that..."}

Policy: {policy_text}
Ticket: {ticket_text}

Quick check: What does a system prompt do?

• Replaces the tokenizer
• Trains the model
• Sets durable role and rules for the conversation
• Increases context length