AI Prompt Engineering Master Guide

Act as an expert prompt engineer with deep knowledge of large language model behaviour, training paradigms, and prompt design patterns across GPT-4, Claude, Gemini, and open-source models. Teach me how to write significantly better prompts for [USE CASE — e.g. "content generation", "data extraction", "code generation", "customer support automation"].

My current AI tool: [ChatGPT / Claude / Gemini / API / Other]
My experience level with prompting: [Beginner / Intermediate / Advanced]
The main task I want to optimise: [DESCRIBE WHAT YOU WANT THE AI TO DO]
Current problem with my prompts: [too generic / too long / wrong format / inconsistent / hallucinations / etc.]

PART 1 — PROMPT ANATOMY

Break down the anatomy of a high-performance prompt. For each element, explain what it does and show a weak vs. strong example:

• Role assignment: how to write a role that actually changes model behaviour (not just "Act as a...") — what makes a role specific enough to be useful
• Context injection: how much context is too much, how to format it, and the "relevant context only" rule
• Task definition: the difference between describing what you want vs. describing the process the model should follow
• Constraints and guardrails: how to prevent the model from doing things you don't want
• Output format specification: how to get consistent, structured outputs every time
• Tone and style direction: how to shape the writing style without over-constraining creativity
• Examples (few-shot): when to include them and how many is optimal
• Chain-of-thought instructions: when "think step by step" works and when it's counterproductive

PART 2 — CORE PROMPTING TECHNIQUES

Explain and demonstrate each technique for [USE CASE]:

ZERO-SHOT PROMPTING:
• When it works: task types where the model doesn't need examples
• The 3 elements a zero-shot prompt must always have
• Example zero-shot prompt for [USE CASE] — bad version and optimised version

FEW-SHOT PROMPTING:
• How many examples is optimal (and why 3 is often better than 10)
• How to format examples for maximum signal
• How to choose which examples to include (diverse, edge-case-representative)
• Write a few-shot prompt template for [USE CASE]

CHAIN-OF-THOUGHT (COT):
• When COT dramatically improves output quality
• How to trigger it without saying "think step by step" (more sophisticated triggers)
• Self-consistency COT: ask for multiple reasoning paths and synthesise
• Write a COT prompt for a complex version of [USE CASE]

ROLE PROMPTING:
• How to write a persona that shapes the model's knowledge access, tone, and reasoning style
• The difference between a shallow role ("Act as a marketer") and a deep role (include background, constraints, goals, perspective)
• Write a deep role definition for the ideal AI assistant for [USE CASE]

CONSTRAINT-FIRST DESIGN:
• Why telling the model what NOT to do is often more effective than what TO do
• How to use negative constraints without accidentally creating new problems
• Write a constraint list for [USE CASE] that prevents the most common failure modes

PART 3 — OUTPUT FORMAT MASTERY

How to get exactly the format you want:
• JSON: how to get clean, parseable JSON every time (including error prevention)
• Markdown: when to use it, how to specify headers, tables, lists exactly
• Tables: how to specify column headers, row types, and data formats
• Numbered lists vs. bullet lists: when each works better
• Length control: how to specify length in terms the model understands (characters, words, sentences, sections)

Write a "format specification block" template I can add to any prompt to lock in the output format.

PART 4 — PROMPT DEBUGGING FRAMEWORK

When your prompt isn't working, follow this diagnostic process:
• Step 1: Identify the failure mode (hallucination / wrong format / too generic / missed instruction / wrong tone / incomplete)
• Step 2: For each failure mode, the most likely cause and the fix
• Step 3: Isolation testing — how to test one variable at a time
• Step 4: The prompt log — how to track and compare prompt versions systematically

PART 5 — ADVANCED PATTERNS

• Prompt chaining: how to break complex tasks into a sequence of smaller, reliable prompts (with a worked example)
• Meta-prompting: using the AI to improve your own prompts — write the meta-prompt template
• Retrieval-augmented prompting: how to inject external knowledge into prompts effectively
• Self-critique pattern: how to make the model critique and improve its own output before you see it
• Constitutional AI approach: how to bake values and rules into prompts for consistent behaviour

PART 6 — PROMPT LIBRARY FOR [USE CASE]

Write 5 ready-to-use, fully optimised prompts for the most common subtasks within [USE CASE]. For each:
• The prompt (ready to copy and use)
• Variables to customise (in [BRACKETS])
• Expected output quality and what to watch for
• When to use this prompt vs. the alternatives