Prompt Engineering

Prompt engineering is the practice of designing instructions and examples that steer model behavior without changing model weights.

You carefully write the request so the model is more likely to respond the way you need.

Prompt design includes instructions, constraints, examples, and output schemas. It interacts with decoding settings (temperature, top-p) and with system prompts in chat stacks.

For many applications, prompt quality and evaluation matter as much as model choice—especially before fine-tuning.

• Assistant behavior shaping
• Structured extraction via schemas
• Evaluation harness prompts

• Fast iteration without retraining
• Improves task adherence
• Enables few-shot patterns

• Brittle across model versions
• Does not fix missing knowledge
• Can be attacked via injection if untrusted text is included

• There is no single magic prompt that works everywhere
• Prompting is not a substitute for security reviews

Is prompt engineering a permanent discipline?

Interfaces will evolve, but specifying goals, constraints, and evals remains necessary regardless of tooling.

• NIST AI RMF

• Brown et al. — Language Models are Few-Shot Learners (in-context learning)