Prompt Engineering

Prompt engineering techniques: give worked examples (few-shot), ask the model to think step-by-step (CoT), define a persona ("You are a senior Python reviewer"), specify output format ("Return JSON with keys..."), and use negative instructions ("Do not include explanations"). Unlike fine-tuning, prompt engineering is free, fast, and reversible. Modern prompting is less mystical than 2022-era "prompt wizardry" because models are better-instructed, but structure still matters.

Effective patterns: (1) Few-shot learning with 3-5 diverse examples, (2) Chain-of-thought triggers for math/logic, (3) Structured output via JSON schema or Pydantic, (4) Self-critique loops ("reflect on your answer"), (5) Decomposition ("break this into sub-tasks"). Anti-patterns: overly long preambles, conflicting instructions, emotional manipulation ("this is urgent!"). Claude's constitutional prompting, OpenAI's structured outputs, and Google's system-instruction API are the 2026 productized approaches. Frontier models in 2026 are robust enough that most prompt "hacks" from 2023 no longer matter.

Prompt engineering as a formal discipline: OpenPrompt, LangChain prompt templates, DSPy for optimized few-shot selection. Few-shot example selection via semantic similarity outperforms random selection by 5-20%. Prompt compression (LLMLingua, H2Olingua) shrinks verbose prompts 2-10× with minimal quality loss. Chain-of-thought triggers lose effectiveness as models become CoT-native (trained to reason without explicit triggers). Structured outputs (JSON schema enforcement at decode time) have largely replaced ad-hoc format prompting.