Prompt Anatomy

Introduction

A production prompt is a structured document — not a sentence. Knowing its parts lets you debug a bad output the way a doctor reads an X-ray: section by section.

Beginner analogy: A prompt is a sandwich. Bread = role + format. Filling = instructions + examples + context. Sauce = constraints. Skip a layer and the whole thing tastes off.

Understanding the topic

Core concepts to understand:

• Role — who the model is ("You are a senior security auditor").
• Task — what to do ("Review this code for OWASP issues").
• Context — background, docs, RAG chunks, prior data.
• Examples — 1–5 input/output pairs that anchor behaviour.
• Constraints — length, tone, what to skip, what to refuse.
• Format — JSON schema, markdown, bullets, table.
• Trigger — "Begin." or the user's actual question.

Syntax reference

Visual workflow / architecture:

┌──── ROLE ─────────────────────────────┐
│ You are a senior security auditor.    │
├──── TASK ─────────────────────────────┤
│ Review the code for OWASP Top 10.     │
├──── CONTEXT ──────────────────────────┤
│ <code snippet>                        │
├──── EXAMPLES ─────────────────────────┤
│ Issue: ...  Fix: ...                  │
├──── CONSTRAINTS ──────────────────────┤
│ Max 5 issues. No marketing language.  │
├──── FORMAT ───────────────────────────┤
│ Return JSON: [{issue, severity, fix}] │
└───────────────────────────────────────┘

Real-world use

Anthropic's official Claude prompting guide is literally this anatomy — role, task, context, examples, format. Every great prompt at every AI company follows the same skeleton.

Best practices

• Always include role + task + format, even for one-off prompts.
• Add examples only when zero-shot fails — they're expensive in tokens.
• Use XML or markdown delimiters to make each section unambiguous.

Common mistakes

• Mixing instructions with examples — model copies the examples literally.
• Skipping the format section and then complaining about messy output.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. List the parts of a production prompt.
• Q2. Which part is the most under-used by beginners?
• Q3. Where would you put RAG context — top, middle or bottom?
• Q4. How do you debug a prompt that returns wrong format?