Prompt Engineering

Zero-shot prompting asks a model to perform a task without providing examples, relying solely on the instruction and the model's pre-trained knowledge. Few-shot prompting provides 1-5+ examples of input-output pairs before the query, helping the model understand the desired format, style, or reasoning pattern. As of 2025, few-shot remains effective for format adherence and domain-specific tasks, though frontier models like GPT-4o and Claude 3.5 Sonnet have improved zero-shot performance. Best practice: use 2-3 diverse examples that cover edge cases; avoid overloading with too many examples which can degrade performance.

Chain-of-thought prompting encourages models to reason through problems step-by-step before arriving at a final answer. This can be triggered explicitly ('Let's think step by step') or implicitly through few-shot examples showing reasoning traces. CoT significantly improves performance on math, logic, and multi-step reasoning tasks. Variants include Zero-shot CoT (just the trigger phrase), Few-shot CoT (examples with reasoning), and Self-Consistency CoT (sampling multiple reasoning paths and voting). Claude models respond well to 'Think through this carefully' triggers; GPT models benefit from explicit 'Step 1, Step 2' formatting.

System prompts define the model's overarching behavior, constraints, and persona across the entire conversation. They have higher priority than user messages in most model architectures (especially Claude). Role prompting assigns a specific identity or expertise ('You are a senior software architect...') to improve domain-specific performance. As of 2025, best practices include: placing core instructions in the system prompt, using explicit constraint lists ('Never reveal internal reasoning'), and defining output formats. Claude's system prompt is set via the 'system' parameter; OpenAI uses 'system' role messages.

Temperature controls randomness in output generation. Values near 0 (e.g., 0.1-0.3) produce deterministic, consistent outputs ideal for factual tasks, code generation, and structured extraction. Higher values (0.7-1.0+) increase creativity and variation, useful for brainstorming, creative writing, and diverse responses. Top-p (nucleus sampling) limits token selection to the most probable tokens summing to probability p; common values are 0.9-0.95. As of 2025, recommended defaults: temperature 0.7 for chat, 0.0-0.2 for deterministic tasks; top-p 0.9-1.0. Some APIs combine temperature and top_p - adjusting both requires care to avoid conflicts.

Context window is the maximum tokens a model can process (prompt + response). As of 2025, windows range from 128K (GPT-4o, GPT-4.1), 200K (Claude 3.5/4), to 1M+ (Gemini 2.0 Pro). Effective prompt engineering requires managing token budgets: system prompts consume tokens permanently; conversation history grows with each turn. Strategies include: summarizing earlier conversation, using RAG to inject only relevant context, and compacting prompts. Token counting varies by model - use tokenizer libraries (tiktoken for OpenAI, Anthropic's tokenizer) for accurate estimates. Overfilling context leads to truncation, which can lose critical instructions.

Prompt injection occurs when user input contains instructions that override or manipulate the model's intended behavior. Attack vectors include: 'Ignore previous instructions and...', hidden instructions in data, and multi-turn manipulation. Defenses include: separating user data from instructions (using delimiters or XML tags), input sanitization, instruction-following hardening in system prompts, and output validation. Claude models have stronger instruction-following defaults; OpenAI models require more explicit system prompt defenses. As of 2025, 'jailbreak' attacks remain an active research area - applications handling untrusted input must treat LLM outputs as untrusted.

Hallucinations are model outputs that seem plausible but are factually incorrect or fabricated. Causes include: insufficient context, ambiguous prompts, and model training on uncertain data. Mitigation strategies: provide explicit source material with instructions to 'only use provided information', request citations or confidence levels, use retrieval-augmented generation (RAG) to ground responses, apply self-consistency checks (ask model to verify its own answer), and prompt for 'I don't know' admissions when uncertain. Claude models tend to be more conservative; GPT models may over-confidently hallucinate. Best practice: never trust factual claims without verification for high-stakes applications.

Multi-turn prompts require maintaining context across conversation turns. Key considerations: conversation history accumulates tokens (potentially exceeding context limits); earlier turns can influence later outputs in unexpected ways; instruction drift can occur as the conversation progresses. Strategies: use system prompts for persistent instructions, summarize or prune older turns, explicitly restate critical constraints in later turns, and use conversation IDs to manage state server-side. As of 2025, Claude has better instruction-following persistence across turns; GPT models may require reminder prompts for complex constraints.