Prompt Engineering

Definition

The practice of crafting effective prompts to get the best results from AI models.

Why It Matters

Closed-weight models don't expose their internals, prompt engineering is the entire surface for tuning behaviour without fine-tuning. The discipline turned from folklore into measurable practice around 2023: structured templates, chain-of-thought, role assignment, output-format constraints, evaluation loops.

Key Points

• Chain-of-thought (CoT): appending 'Let's think step by step' improves multi-step math and logic accuracy by 10–30 percentage points on the same model.
• ReAct (Reason + Act) interleaves reasoning traces with tool calls, the backbone architecture behind most agentic frameworks.
• Output format locking: specify the JSON schema or output structure before describing the task for more reliable parsable output.
• Structured elicitation: asking the model to list pros/cons, alternatives, or counterarguments before concluding reduces overconfident single-option answers.
• Evaluation-driven prompting: measure output quality on a sample of real inputs before finalising a prompt, intuition about what 'should work' is often wrong.

Example

Adding "Let's think step by step" before a math problem improves accuracy on multi-step problems by 10–30 percentage points on the same model. Asking for JSON output before describing the schema is more reliable than describing the schema first. Both count as prompt engineering.

Related Terms

• PromptThe input text you give to an AI model. Better prompts lead to better outputs.
• Few-Shot LearningGiving an AI model a few examples in the prompt to guide its output.
• Zero-Shot LearningAn AI model performing a task without any specific examples, just from its general training.

Quick Facts