Constraints in Prompts

Introduction

Constraints tell the model what NOT to do, how long to be, what to refuse, what voice to avoid. They're the guardrails that keep output usable.

Beginner analogy: Like dietary restrictions on a recipe — 'no nuts, gluten-free, under 20 minutes'. Constraints shape the dish without prescribing it.

Understanding the topic

Core concepts to understand:

• Length: 'under 100 words', 'exactly 3 bullets'.
• Tone: 'professional, no emojis', 'casual, friendly'.
• Content: 'no medical advice', 'no personal opinion'.
• Format: 'no markdown', 'JSON only'.
• Refusals: 'if unsure, reply EXACTLY: I cannot answer'.

Syntax reference

Visual workflow / architecture:

Constraints checklist:
  ☑ length      ≤100 words
  ☑ tone        professional
  ☑ content     no claims about effectiveness
  ☑ format      JSON, no comments
  ☑ refusal     say "I don't know" if unsure

Real-world use

Legal, medical and financial AI assistants use heavy constraint sections to stay compliant.

Best practices

• Constrain less than you think — too many constraints reduce quality.
• Always include a refusal pattern.
• Test constraints with adversarial inputs.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why are constraints important in production?
• Q2. Where do constraints go in the prompt?
• Q3. How can constraints reduce hallucinations?
• Q4. How do you test that constraints hold?