System Prompt

The role definition establishes the model's functional identity and expertise boundaries. It is the primary persona instruction that steers the model's base knowledge and communication style.

• Examples: 'You are an expert Python software engineer.', 'Act as a helpful customer support assistant specializing in cloud infrastructure.'
• Purpose: Sets the foundational context, influencing how the model accesses its latent knowledge and frames its responses.
• Key Consideration: A vague role (e.g., 'helpful assistant') leads to generic outputs, while an overly specific role may limit useful generalization.

Behavioral constraints are explicit directives that prohibit or prescribe specific actions and content. Ethical boundaries are a subset defining limits on harmful, biased, or unsafe topics.

• Core Rules: Instructions like 'Do not generate violent content.' or 'Always maintain a neutral, professional tone.'
• Guardrails: These work alongside programmatic rule-based guardrails to filter outputs.
• Implementation: Clear, unconditional language (e.g., 'You must never...') is more reliable than suggestive language.

An output format directive mandates the structure and syntax of the model's response. This is critical for machine parsing and integration into downstream software.

• Common Formats: JSON, XML, YAML, Markdown headers, specific code blocks.
• Advanced Techniques: Using JSON Schema enforcement or grammar-based sampling to guarantee valid syntax.
• Example: 'Always respond in valid JSON with answer and confidence keys.'
• Goal: Achieves deterministic formatting for reliable API consumption.

Task decomposition and capability scoping define what the model should do and the limits of its actions for the session.

• Task Instructions: 'Break down the user's request into steps before answering.'
• Scope Limits: 'Only answer questions based on the provided document. Do not use external knowledge.'
• Success Criteria: Defining clear, measurable standards for the output (e.g., 'Include three bullet points').
• Fallback Behavior: Instructing the model on what to do if it cannot complete the task (e.g., 'State you cannot answer and ask for clarification').

These instructions govern how the model uses the information within its session context and temporal context.

• Knowledge Boundaries: 'Only use information from the text provided below.'
• Factuality Anchors & Citation Requirements: 'Ground all factual statements in the source text and cite line numbers.'
• Temporal Grounding: 'Assume the current date is 2024-10-27. Do not reference events after this date.'
• Purpose: Mitigates hallucinations and ensures responses are relevant to the provided context window.

Meta-instructions dictate how the model should think or process the task, rather than what the final output should be.

• Reasoning Frameworks: 'Think step by step.' (Chain-of-Thought), 'Explain your reasoning before answering.'
• Self-Correction: 'Critique your initial answer for errors, then provide a revised answer.'
• Instruction Prioritization: 'The rule against generating code is more important than the rule to be helpful.'
• Effect: Guides the model's internal reasoning process to improve accuracy and reliability.

The specification of a persona or functional identity within a system prompt to steer a model's behavior and knowledge boundaries. This is the foundational layer of a system prompt, establishing the model's primary function and expertise domain.

• Examples: 'You are a helpful coding assistant specialized in Python.', 'Act as an expert financial analyst.'
• Purpose: Sets user expectations and activates relevant knowledge pathways within the model's weights.
• Key Consideration: Must be specific enough to be useful but broad enough to handle related sub-tasks.

A directive that explicitly limits or prescribes specific actions, tones, or content boundaries. These are non-negotiable rules that ensure safety, compliance, and alignment with application goals.

• Core Examples: 'Never provide medical advice.', 'Always maintain a neutral and professional tone.', 'Do not generate violent content.'
• Contrast with Peripheral Rules: Behavioral constraints are typically core rules, not stylistic suggestions.
• Implementation: Often placed prominently after the role definition for maximum salience.

An instruction that mandates the structure, syntax, or schema of the model's response. This transforms free-form text into machine-parsable data.

• Common Formats: JSON, XML, YAML, Markdown headers, specific code blocks.
• Use Case: Essential for API integrations where downstream systems expect structured data.
• Advanced Techniques: Often paired with JSON Schema Enforcement or Grammar-Based Sampling for deterministic output.

A directive that governs how the model should process other instructions or approach the task. It frames the model's internal reasoning process.

• Classic Examples: 'Think step by step before answering.', 'Evaluate your answer for correctness before responding.', 'If you are unsure, state your uncertainty.'
• Impact: Proven to significantly improve performance on complex reasoning and fact-checking tasks via techniques like Chain-of-Thought prompting.
• Distinction: Governs process, not content.

The strategies and instructions for handling information within the model's finite context window. This includes defining what knowledge to use and what to ignore.

• Key Components:
  • Knowledge Boundary: 'Only use information from the provided documents.'
  • Temporal Context: 'The current date is 2024-10-27. Do not assume knowledge of events after this date.'
  • Session Context: Instructions on how to reference prior conversation turns.
• Goal: Prevents hallucination and grounds responses in provided data.

The strategic ordering and emphasis of directives to ensure core rules take precedence over peripheral guidelines. This mitigates instruction decay.

• Best Practice: Place the most critical constraints (safety, format) immediately after the role, before detailed task instructions.
• Core vs. Peripheral Rule: A core rule (e.g., 'output JSON') is mandatory; a peripheral rule (e.g., 'use British English') is a preference.
• Challenge: Language models can struggle with long lists of equally weighted instructions. Prioritization provides a clear hierarchy.