Basic Prompting Techniques

Few-shot prompting involves providing multiple examples of input-output pairs that demonstrate how to perform a task, enabling models to learn patterns from examples rather than requiring explicit instruction. This topic covers in-context learning, the mechanism by which LLMs adapt to demonstrated patterns without any weight updates. Few-shot prompting is powerful...

Few-shot prompting involves providing multiple examples of input-output pairs that demonstrate how to perform a task, enabling models to learn patterns from examples rather than requiring explicit instruction. This topic covers in-context learning, the mechanism by which LLMs adapt to demonstrated patterns without any weight updates. Few-shot prompting is powerful for tasks where the pattern is complex, where desired output format is non-standard, or where precise behavior needs to be demonstrated. The topic explains how to select high-quality examples including choosing examples that cover diversity of input types, represent the full range of desired behaviors, and are realistic rather than contrived. The topic addresses the quantity-quality tradeoff, noting that a few well-chosen diverse examples often outperform many redundant examples. Example ordering is discussed, with evidence suggesting that the order of examples can affect performance. The topic covers techniques for constructing examples including creating synthetic examples when real ones aren't available, annotating examples with explanations, and using consistent formatting across examples. The topic discusses how example selection interacts with task difficulty, with more complex tasks potentially benefiting from more or more detailed examples. The topic addresses computational costs of few-shot prompting, noting the token overhead when including examples, and discussing strategies for compressing examples without losing information.

Output formatting techniques specify how the model should structure its response, enabling both human readability and machine parsing. Specifying format is critical for production systems where outputs feed into other processes. Common formats include JSON for structured data, bullet points for concise lists, tables for comparative information, and prose for...

Output formatting techniques specify how the model should structure its response, enabling both human readability and machine parsing. Specifying format is critical for production systems where outputs feed into other processes. Common formats include JSON for structured data, bullet points for concise lists, tables for comparative information, and prose for narrative content. The topic explains how explicit format specifications improve output consistency compared to implicit expectations. The topic covers techniques for specifying formats including providing the desired schema, giving an example of properly formatted output, and explicitly stating format requirements. The topic discusses format negotiation, recognizing that some formats are more natural for models than others and that complex custom formats may require combination with examples. The topic addresses constraint specification including length limits, forbidden words, and required elements. The topic covers evaluation of whether specified formats are achievable given model capabilities, noting that some requests conflict with model properties (e.g., requesting always-true certainty when models express uncertainty). The topic includes discussion of common pitfalls such as requesting formats that encourage hallucination or requesting outputs that exceed context window limits.

Zero-shot prompting represents the simplest approach to using LLMs, where a model is asked to perform a task based purely on its pre-trained knowledge and the instruction provided, without seeing examples of successful task completion. This topic explains how modern LLMs acquire capabilities across diverse tasks during pre-training, enabling zero-shot...

Zero-shot prompting represents the simplest approach to using LLMs, where a model is asked to perform a task based purely on its pre-trained knowledge and the instruction provided, without seeing examples of successful task completion. This topic explains how modern LLMs acquire capabilities across diverse tasks during pre-training, enabling zero-shot performance. Zero-shot prompting is most effective for well-defined tasks where models possess training knowledge, such as summarization, translation, or question answering on widely-known topics. The topic covers best practices for zero-shot prompting including clear task definition, specification of constraints, and explicit output format requests. Zero-shot prompts should be structured clearly with distinct sections for instruction, context, and input. The topic explains why zero-shot works well for some tasks and poorly for others, with performance depending on whether the model has encountered similar content during training and whether the task is well-specified. The topic addresses techniques for improving zero-shot performance including more specific instructions, example-based explanation of concepts, and specification of expected output characteristics. Zero-shot prompting is positioned as the starting point for prompt development, with more advanced techniques applied when zero-shot performance is insufficient.

One-shot prompting provides a single example of the task to be performed, which helps guide the model toward the desired behavior without the overhead of multiple examples. This topic explains when one example is sufficient, typically for well-defined patterns where a single example clearly illustrates the expected transformation. One-shot examples...

One-shot prompting provides a single example of the task to be performed, which helps guide the model toward the desired behavior without the overhead of multiple examples. This topic explains when one example is sufficient, typically for well-defined patterns where a single example clearly illustrates the expected transformation. One-shot examples are most effective when the pattern is obvious and consistent. The topic covers designing effective examples that clearly separate input from output, use realistic content, and demonstrate the exact format and style desired. The topic explains how examples should be clearly distinguished from the actual task input, using visual separators or explicit labeling. The topic discusses integration of one-shot examples with other prompt elements, typically placing examples between the instruction and the actual input to be processed. The topic addresses variation across tasks, noting that some tasks require more complex examples while others benefit from simpler illustrations. The topic includes comparative analysis of when one-shot is superior to zero-shot (when format matters) versus when few-shot is needed (when pattern complexity increases).

Instruction-based prompting emphasizes the use of clear, explicit task descriptions to guide LLM behavior. This approach works best when the task can be specified through direct instructions without relying on examples. Effective instructions include explicit task description that defines what the model should do, constraints that specify what should and...

Instruction-based prompting emphasizes the use of clear, explicit task descriptions to guide LLM behavior. This approach works best when the task can be specified through direct instructions without relying on examples. Effective instructions include explicit task description that defines what the model should do, constraints that specify what should and shouldn't be done, expected output format, and success criteria that clarify when the task has been completed successfully. The topic covers techniques for writing instruction-based prompts including use of specific verbs that indicate precise actions, separation of distinct requirements, and explicit rather than implicit specifications. The topic discusses how instruction specificity must balance between clarity and flexibility, with over-specific instructions sometimes preventing creative problem-solving. The topic addresses how instructions should be tailored to the audience, recognizing that LLMs process language differently than humans. The topic includes analysis of instruction structure including optimal ordering where general task description precedes constraints, followed by detailed requirements. The topic covers iteration strategies for improving instructions based on observed outputs, including diagnoses of instruction inadequacy and targeted improvements.

Role-based prompting assigns a persona or role to the LLM, establishing context that influences how the model approaches tasks. This technique is powerful for tone control, domain expertise simulation, and behavioral alignment. A role might be 'You are a professional copywriter' to emphasize marketing language, 'You are a skeptical analyst'...

Role-based prompting assigns a persona or role to the LLM, establishing context that influences how the model approaches tasks. This technique is powerful for tone control, domain expertise simulation, and behavioral alignment. A role might be 'You are a professional copywriter' to emphasize marketing language, 'You are a skeptical analyst' to encourage critical evaluation, or 'You are a Python expert' to focus on technical accuracy. The topic explains how roles activate relevant knowledge from training without requiring explicit instruction of that knowledge. The topic covers selection of roles that match the task requirements and desired output characteristics, noting that generic roles like 'helpful assistant' activate different behaviors than specific roles like 'Stanford professor of AI ethics.' Role-based prompting can control tone by assigning roles with associated linguistic patterns, such as 'formal regulatory compliance expert' versus 'casual tech blogger.' The topic addresses combination of roles with other prompting techniques, such as role assignment combined with few-shot examples. The topic discusses limitations of role-based prompting, noting that roles are suggestions rather than constraints and that models may not perfectly adopt assigned personas. The topic covers system messages as a architectural approach to role specification, used in API-based systems to establish model behavior consistently across conversation.