Conditional Instruction

The core mechanism of a conditional instruction. It directs the model to evaluate a condition in the user input and execute a corresponding action.

• Condition: A testable attribute of the input (e.g., if the user asks for code, if the query contains a date).
• Action: The prescribed model behavior if the condition is met (e.g., then respond in Python, then format the answer as a timeline).

This creates deterministic, context-aware responses without requiring multiple, separate prompts.

An extension of if-then logic for handling multiple, mutually exclusive scenarios. It instructs the model to select one behavior from a set based on a key input characteristic.

Example Directive: Based on the user's question type, format your answer as follows: - If it's a definition: Provide a concise, dictionary-style entry. - If it's a comparison: Use a table with columns for 'Feature A' and 'Feature B'. - If it's a step-by-step guide: Use a numbered list.

This is more efficient and less prone to conflict than a chain of separate if statements.

Conditionals activate specific instructions only when relevant context is present, preventing instruction overload and instruction decay. The model ignores branches that don't apply to the current query.

Key Benefit: It allows a single, comprehensive system prompt to safely contain rules for diverse tasks without those rules interfering with each other. For instance, a coding assistant's prompt can have strict security rules for shell commands that are only invoked if the user's request involves system operations.

The condition often requires the model to first classify the user input. This implicit step is a foundational part of the conditional's execution.

Common Classification Triggers:

• Intent: Is the user asking, commanding, or clarifying?
• Domain: Does the query relate to code, science, or creative writing?
• Complexity: Is the request simple or multi-faceted?
• Format Request: Did the user ask for a list, JSON, or a summary?

The conditional instruction effectively embeds a lightweight classification task before generating the final output.

The primary engineering goal. By mapping specific input patterns to specific output formats and behaviors, conditional instructions route the model's generative process down a predefined, verifiable path.

This reduces variance and increases reliability, making the model's behavior more predictable and suitable for integration into software pipelines. It is a foundational technique for implementing structured generation and response schemas dynamically, based on the query itself.

A critical component for robustness. A well-designed conditional instruction includes a default or else clause to define the model's fallback behavior when no explicit condition is met.

Example: If the user's request matches one of the formats above, use it. Otherwise, provide a clear, well-structured paragraph and ask if a specific format would be helpful.

This clause handles edge cases, ambiguous inputs, and novel queries, ensuring the system remains helpful and does not fail silently. It directly implements an error handling directive.

This pattern directs the model to analyze the user's query and select a specific response protocol.

• If the user asks for a factual summary, then provide a bulleted list of key points.
• If the user asks for creative writing, then generate a narrative in the first person.
• If the user's query contains technical jargon, then assume an expert audience and do not simplify terms.

Example Prompt Snippet: Analyze the user's request. IF it is a request for data, respond with a table. IF it is a request for analysis, respond with a paragraph.

Conditionals dynamically assign a model's persona based on context clues in the conversation.

• If the user mentions 'legal advice', then adopt the persona of a legal consultant and preface responses with 'I am not a lawyer, but based on general principles...'
• If the user's previous message was in Spanish, then switch to responding entirely in Spanish.
• If the query begins with 'Debug this code:', then assume the role of a senior software engineer.

This creates a multi-role assistant within a single system prompt, eliminating the need for separate model sessions.

This structure mandates a primary output format but provides a fallback instruction if the model cannot comply, ensuring robust operation.

• Primary Instruction: Always output your final answer as a valid JSON object with 'answer' and 'confidence' fields.
• Conditional Fallback: IF you cannot structure the answer as JSON, THEN output 'ERROR: Cannot format as JSON' and provide the answer in plain text.

This is critical for downstream API integration where a parsing failure must be handled gracefully without breaking the application workflow.

Explicit conditional logic acts as a first-layer filter for content moderation and ethical boundaries.

• If the user asks for instructions on illegal activities, then respond with: I cannot provide guidance on that topic.
• If the query requests medical diagnosis, then respond with: I am an AI and cannot provide medical advice. Please consult a healthcare professional.
• If the input contains profanity, then respond neutrally without echoing the language.

These are core rules that take precedence over all other instructions in the prompt, implementing a form of Constitutional AI at the prompt level.

Instructions adjust the verbosity and technical depth of responses based on implicit signals about the user.

• If the user says 'explain like I'm 5', then use simple analogies and avoid jargon.
• If the user provides a complex, multi-part query, then structure the response with clear headings and a summary.
• If the conversation history shows repeated requests for clarification, then increase the detail in subsequent answers.

This pattern enables audience adaptation without requiring the user to explicitly re-state their preference.

Conditionals manage the model's use of time-sensitive or provided information to prevent hallucinations.

• If the user asks about current events, then state your knowledge cutoff date and do not speculate about events after it.
• If a 'source_text' is provided in the context, then base all factual answers solely on that text and cite relevant passages.
• If the user asks 'based on the document above...', then restrict your reasoning to the preceding context, ignoring general knowledge.

This creates a knowledge boundary and acts as a factuality anchor, which is foundational for Retrieval-Augmented Generation (RAG) systems.

A system prompt is the foundational, high-level instruction provided at the start of a session that defines a model's role, behavior, and constraints for all subsequent interactions. It establishes the operational framework within which conditional instructions and other directives are executed.

• Primary Purpose: Sets the global context and rules for the interaction.
• Relationship to Conditional Instructions: Serves as the container where conditional logic is typically embedded.
• Example: "You are a financial analyst. Always cite sources. If the user asks for a forecast, provide high, medium, and low scenarios."

A meta-instruction is a directive that governs how the model should process other instructions or perform internal reasoning. It often precedes or wraps conditional logic to ensure proper execution.

• Core Function: Controls the model's cognitive or procedural approach.
• Common Examples: "Think step by step before answering.", "Evaluate the user's query to determine its intent first.", "Critique your own answer for factual accuracy."
• Use with Conditionals: A meta-instruction like "First, classify the user's request." is frequently the trigger for subsequent conditional branches.

A task decomposition prompt instructs the model to break a complex request into a sequence of simpler, executable steps. This is a higher-order form of conditional logic where the branching is determined and executed by the model itself.

• Mechanism: Directs the model to create its own execution plan based on input analysis.
• Contrast with Conditional Instruction: While a conditional instruction provides predefined if-then branches, a task decomposition prompt elicits the model to generate the branches dynamically.
• Example: "For this complex project plan request, first outline the major phases, then detail the tasks for each phase, and finally identify potential risks."

A fallback behavior directive is a specific type of conditional instruction that defines the model's response when it cannot successfully execute the primary request or when input conditions are not met. It is essential for robust error handling.

• Purpose: Ensures graceful failure and maintains user experience.
• Typical Triggers: Ambiguity, out-of-scope requests, missing information, or internal uncertainty.
• Standard Structure: "If you cannot [perform primary task] because [condition], then [specified fallback action]."
• Example: "If the user's question requires knowledge after October 2023, politely state your knowledge cutoff and ask if they have a more recent source."

Structured output generation refers to techniques that force a model's response into a precise, machine-readable format like JSON, XML, or YAML. Conditional instructions are often used to determine which schema or structure to apply based on the input.

• Key Link: The "then" clause of a conditional instruction frequently mandates a specific output schema.
• Enforcement Methods: Often combined with JSON Schema enforcement or grammar-based sampling.
• Conditional Example: "If the user asks for weather data, output JSON with {city, temp_c, condition}. If they ask for a news summary, output a markdown list."

Instruction prioritization is the strategic ordering and weighting of directives within a system prompt to manage conflicts and ensure core rules are followed. It provides a hierarchy that conditional instructions must operate within.

• Core vs. Peripheral Rules: Establishes which instructions are non-negotiable (core) and which are flexible guidelines (peripheral).
• Impact on Conditionals: A conditional instruction cannot override a higher-priority core rule (e.g., a conditional to "provide medical advice" would be invalid if a core rule states "never provide medical advice").
• Design Principle: Critical for preventing logical contradictions in complex prompt architectures containing multiple conditionals.