The rise of Large Language Models (LLMs) has brought exciting possibilities for human-computer interaction. However, to fully unlock the potential of these powerful AI models, one crucial skill is required: prompt engineering.

This rapidly evolving field focuses on crafting effective prompts that leverage the capabilities of LLMs, enabling them to understand intent, follow instructions, and generate desired outputs. As AI becomes increasingly embedded in various applications, prompt engineering plays a critical role in ensuring accurate, relevant, and safe interactions.

In AI dialogue, system, user, and assistant are the core interaction roles, while tool/function serves as an extension that enhances the assistant's capabilities.

1. System: The Invisible Rule Setter

The system role consists of instructions preset by developers. These are not exposed to users directly but instead guide the assistant, defining boundaries, scope, tone, and style—essentially serving as the AI's instruction manual.

Core Functions:

• Define identity and goals
  Example: "You are a professional math tutor. Explain junior-high math step by step, and avoid advanced math terminology."
• Set interaction rules
  Example: "If a user asks non-math questions, politely redirect to math. Keep answers under 300 words and emphasize formula derivation."
• Provide background information
  Example: "Here is a company's product specification sheet. When asked about products, only use this sheet and do not invent information."
• Define output format
  Example: "When using the calculator, respond strictly in the format: [Tool: Calculator, Parameters: {expression: '2+34', precision: integer}]"

2. User: The Origin of Demand

The user role represents the source of interaction—the user expresses needs through text, voice, images, or other forms. User inputs serve as the primary basis for the assistant's responses.

Core Functions:

• Initiate requests
  Example: "Calculate the number of working days in May 2024" / "Explain the core principles of relativity" / "Write a product marketing copy."
• Provide additional details
  Example: Assistant asks: "Is your product in electronics or cosmetics?" User: "Cosmetics—lipstick."
• Feedback and adjustment
  Example: "The copy sounds too formal. Can you make it more playful?" / "The calculation is wrong, check again."

3. Assistant: The Executor and Bridge

The assistant is the AI system's visible role. Its core duties are to interpret system rules + parse user intent + produce compliant responses. In systems with tools, it also decides whether and how to invoke them.

Core Functions:

• Understand intent
  Example: User asks "Should I bring an umbrella tomorrow?" → Assistant interprets as "Check tomorrow's rainfall probability."
• Directly respond
  Example: "The core principle of relativity is…" / "2 + 3 = 5."
• Decide tool use
  Example: "Call weather API for Beijing's forecast" / "Call calculator for 12345 × 6789."
• Integrate tool results
  Example: Tool returns: "Beijing May 1, 2024: Rain probability 60%." Assistant says: "Beijing's rain chance is 60%. We suggest carrying an umbrella."

4. Extended Role: Tool/Function – Assistant's Power Extension

Tools and Functions handle tasks that the assistant cannot perform on its own. They are essentially callable programs, APIs, or services, e.g., weather APIs, calculators, document parsers, or map services.

Without tools, assistants rely solely on training data. With tools, they can access real-time info, perform complex operations, and support far broader use cases.

Core Functions:

• Fill capability gaps: Handle tasks beyond model limitations.
• Improve precision: Avoid outdated/guessed answers (e.g., get live stock prices instead of relying on old data).
• Reduce user effort: Users state their needs, the Assistant handles tool calls automatically.

User asks for the PM2.5 index of Santa Clara, California on May 1, 2024, and requests an analysis report.

Workflow:

1. System preset rule: "Must call Air Quality API; report must include index value + grade + health advice, and provide a download link."
2. User initiates: the query and later receives the result.
3. Assistant acts as decision-maker: decides to call both the Air Quality API and a PDF report generator.
4. Tool executes: returns live data and generates the report.

• Improved model performance: Clear prompts yield more accurate, relevant, and informative responses.
• Bias & harm reduction: Well-crafted prompts help reduce biased or unsafe outputs, when combined with model-level safeguards.
• Greater control & predictability: Prompts guide model behavior toward consistent, expected results.
• Enhanced user experience: Clear prompts enable smoother, more intuitive interactions.

1. Define clear goals

2. Provide context & background

3. Use few-shot prompting

4. Use chain-of-thought prompting