System Prompt

This component explicitly assigns an identity and expertise to the model, such as 'You are a senior software architect' or 'Act as a helpful customer support agent.' It sets the tone, domain knowledge, and communication style, priming the model's internal representations to align with the specified character. This is a form of role prompting that significantly narrows the model's response distribution.

• Example: You are an expert financial analyst specializing in renewable energy markets. Provide concise, data-driven insights.

This is the core operational instruction that defines the primary objective the model must accomplish. It should be clear, unambiguous, and action-oriented. Effective task specification often uses imperative verbs and outlines the expected output's purpose.

• Example: Your task is to analyze the provided user query and a set of retrieved documents. Synthesize a comprehensive answer that cites specific facts from the documents. This directly enables patterns like Retrieval-Augmented Generation (RAG).

This component establishes hard boundaries for the model's behavior to ensure safety, compliance, and operational correctness. It proactively mitigates risks like hallucination and prompt injection.

• Key constraints include:
  • Output Format: Always respond in valid JSON with the keys 'summary' and 'confidence_score'.
  • Content Boundaries: Do not generate creative fiction. Do not provide medical or legal advice.
  • Source Grounding: Only use information present in the provided context. If the answer is unknown, state 'I cannot find that information.'

This defines the information boundary for the model's response. It instructs the model on which data sources it should prioritize or ignore, which is critical for factual accuracy. This component is essential for managing the context window effectively.

• Examples:
  • For a chatbot: Use only the company knowledge base documents provided below.
  • For a coding assistant: Refer to the official Python 3.11 documentation and the user's code snippet.
  • It explicitly separates system instructions from retrievable data in RAG architectures.

This component instructs the model how to think, not just what to output. It scaffolds the internal reasoning process to improve reliability on complex tasks.

• Common techniques integrated here:
  • Chain-of-Thought (CoT): Think through this problem step by step before giving your final answer.
  • Self-Consistency: Generate three possible solutions, then select the best one.
  • ReAct (Reason + Act): First, reason about what information is needed. Then, decide if a tool call is required. This turns the system prompt into a lightweight reasoning architecture.

This defines the rules of engagement for multi-turn dialogues or interactions with external systems. It manages conversational state and tool use.

• Key elements:
  • Turn Management: Maintain a concise history of the last 5 exchanges.
  • Tool Calling / Function Calling: When you need current data, output a JSON object with the 'tool_call' key.
  • Clarification Protocol: If the user request is ambiguous, ask exactly one clarifying question. This component is vital for building robust, stateful agentic applications.

This archetype explicitly assigns a persona and domain expertise to the model, constraining its responses to a specific professional context.

• Core Instruction: You are an expert financial analyst with 15 years of experience at a major investment bank.
• Key Behaviors: Responses adopt a formal, data-driven tone, use financial terminology precisely, and prioritize quantitative reasoning over speculative narratives.
• Use Case: Automating the generation of earnings report summaries, risk assessments, or investment memos where consistent professional voice is critical.

This prompt prioritizes structural conformity, instructing the model to produce responses in a strict, machine-parsable schema.

• Core Instruction: Always output your final answer as a valid JSON object with the exact keys: 'summary', 'confidence_score', 'key_entities'.
• Key Behaviors: The model suppresses natural language explanations unless explicitly requested, validates its own output against the schema, and uses delimiters like triple backticks (```json).
• Use Case: Integrating LLM outputs directly into downstream software pipelines, APIs, or data lakes where consistent structure is non-negotiable.

This archetype establishes immutable ethical and operational boundaries, often layered on top of other instructions.

• Core Instruction: Under no circumstances provide medical diagnosis advice, financial guarantees, or legally binding statements. If a query requests this, decline and explain the limitation.
• Key Behaviors: The model performs a pre-response compliance check, uses hedging language for sensitive topics, and has a defined refusal protocol for out-of-bounds requests.
• Use Case: Customer-facing chatbots, content moderation assistants, or any public interface where mitigating liability and adhering to regulations (like the EU AI Act) is paramount.

This prompt defines a step-by-step reasoning and action loop, often integrating Chain-of-Thought (CoT) with tool-calling capabilities.

• Core Instruction: Follow this process: 1. Reason about the user's goal. 2. If needed, use the provided get_weather tool. 3. Synthesize the information. 4. Provide a final, concise answer.
• Key Behaviors: The model explicitly verbalizes its internal reasoning, structures its response into defined phases, and formats tool calls according to a specified protocol (e.g., JSON).
• Use Case: Building autonomous agents for customer support, data analysis workflows, or operational research where transparent, multi-step problem-solving is required.

This archetype explicitly governs the model's interaction with retrieved documents in a Retrieval-Augmented Generation (RAG) system, combating hallucination.

• Core Instruction: Your knowledge is strictly limited to the provided context documents. Base your answer solely on them. If the answer is not found, say 'I cannot find that information in the provided documents.'
• Key Behaviors: The model cites specific document passages, avoids extrapolating beyond the provided text, and has a clear fallback behavior for information gaps.
• Use Case: Enterprise knowledge assistants, technical support bots, or legal document review systems where factual accuracy and source grounding are essential.

This prompt uses stylistic and tonal constraints to shape generative output for branding or creative consistency.

• Core Instruction: Write in the style of a classic hard-boiled detective novel: use short, punchy sentences, vivid metaphors, and a cynical first-person perspective.
• Key Behaviors: The model adheres to specific lexical choices, narrative pacing, and emotional tone, effectively acting as a brand voice simulator.
• Use Case: Marketing content generation, script writing for games or media, and automated product description writing where maintaining a specific creative voice at scale is necessary.

A prompting technique that instructs an LLM to adopt a specific persona, expertise, or perspective (e.g., 'You are a senior software architect reviewing this code'). This is a fundamental application of a system prompt, used to tailor tone, domain knowledge, and response format. It leverages the model's pre-trained understanding of roles to constrain and guide its output.

• Example: A system prompt stating You are a helpful customer support agent will generate more empathetic and solution-oriented responses than a generic assistant.

A higher-order prompt that instructs an LLM to generate, analyze, or refine another prompt. It is used in automated prompt engineering workflows. A meta-prompt itself is a type of system prompt, but its goal is prompt creation rather than direct task execution.

• Use Case: You are an expert prompt engineer. Given the following task description, generate an optimal system prompt for a language model to accomplish it.
• This enables iterative optimization and the generation of task-specific system instructions programmatically.

A reusable, parameterized blueprint for constructing prompts, containing static instructions (often the system prompt), variables (placeholders), and a structured format. It ensures consistency, efficiency, and version control in production applications.

• Components: A template typically includes a system role definition, a user message slot, and sometimes few-shot examples and output format specifications.
• Engineering Benefit: Separates application logic from prompt text, allowing updates to the system instruction without code changes.

A supervised fine-tuning (SFT) process where a base LLM is trained on a dataset of (instruction, output) pairs. This teaches the model to better understand and follow natural language commands, making it more responsive to system prompts and user instructions.

• Direct Relationship: The quality of a system prompt's execution is heavily dependent on the model's underlying instruction-following capability, which is honed during instruction tuning.
• Models like GPT-4 and Claude undergo extensive instruction tuning to become effective 'assistants' that reliably adhere to provided system guidelines.

A paradigm (Reasoning + Acting) that combines chain-of-thought reasoning with actionable steps (tool/API calls) within a single prompt. The system prompt in a ReAct setup must define the agent's role, available tools, and the required format for interleaving thoughts and actions.

• Structure: The system instruction typically outlines a loop: Thought: [reason about current step], Action: [call tool X], Observation: [tool result].
• This demonstrates how a complex system prompt orchestrates not just style, but a multi-step cognitive process involving external systems.