Prompt Engineering Patterns

Improve output reliability with practical prompting techniques

Overview

Prompt quality has the biggest impact on response quality. Use consistent structures to improve reliability.

Core Patterns

1) Role + Task + Output Format

You are a support analyst.
Task: Summarize the incident.
Output: JSON with keys root_cause, impact, next_step.

2) Constraint-Driven Prompting

Add explicit limits:

• Word count ceiling
• Required sections
• Forbidden content

3) Few-Shot Examples

Provide 1-2 examples when format consistency is critical.

4) Self-Check Instruction

Ask for final verification against constraints before completion.

Evaluation Loop

• Start minimal.
• Add one constraint at a time.
• Retry only after recording what changed.

When benchmarking models, keep the prompt static and adjust only generation parameters per trial.