📋 Prompt Engineering Patterns — Quick Reference Cheatsheet

Use this as a scannable reference when you know what you need but want to quickly find the right pattern.

🧱 Foundational Patterns

#	Pattern	One-Line Description	When To Use	Difficulty
🔗 01	Chain-of-Thought	Force the model to show its reasoning before concluding	Multi-step reasoning, math, complex analysis	Beginner
🎯 02	Few-Shot Examples	Show 2-3 examples of desired output to set quality bar	Consistent tone, format, or style across outputs	Beginner
🎭 03	Role Assignment	Give the model a specific expert persona and behavioral traits	Need domain expertise or specific communication style	Beginner
📋 04	Output Formatting	Define exact schema/structure for model output	Parsing outputs programmatically or comparing results	Beginner
🪜 05	Step-by-Step Decomposition	Break complex task into sequential subtasks in one prompt	Task too complex for one pass but doesn't need a pipeline	Beginner

#	Pattern	One-Line Description	When To Use	Difficulty
🔄 06	Self-Consistency Checking	Analyze from multiple frameworks, then synthesize	High-stakes decisions where wrong answers have consequences	Intermediate
🚧 07	Constraint Setting	Add explicit boundaries on length, scope, tone, and content	Model goes off-track in predictable, repeatable ways	Intermediate
🚫 08	Negative Examples	Show what NOT to do with annotated bad examples	Persistent failure modes that instructions alone can't fix	Intermediate
📐 09	Context Window Management	Structure long inputs to prevent "lost in the middle" failures	Working with large documents or many data points	Intermediate
🌡️ 10	Temperature & Sampling	Match randomness parameters to task type	Outputs too repetitive (temp too low) or chaotic (too high)	Intermediate

#	Pattern	One-Line Description	When To Use	Difficulty
🏗️ 11	System Prompt Architecture	Multi-section system prompt with identity, capabilities, boundaries	Building any persistent AI application or chatbot	Advanced
🛡️ 12	Adversarial Guardrails	Defense-in-depth against prompt injection and misuse	User-facing prompts with untrusted input	Advanced
✅ 13	Chain of Verification	Model verifies its own claims against source material	Factual accuracy matters and hallucination is a real risk	Advanced
⛓️ 14	Prompt Chaining Pipelines	Multi-prompt pipeline with code logic between stages	Task needs different models, validation, or branching	Advanced
🎲 15	Dynamic Few-Shot Selection	Programmatically choose examples based on input similarity	Large example bank, limited context, varying input types	Advanced

#	Pattern	One-Line Description	When To Use	Difficulty
📊 16	Evaluation-Driven Iteration	Systematic scoring framework replacing gut-feel tuning	Moving from "works in testing" to "works at scale"	Advanced
🧪 17	Regression Testing	Automated test suite that catches prompt regressions	Prompts in production where changes could break things	Advanced
🔀 18	Multi-Model Routing	Route requests to optimal model by complexity and cost	Balancing quality, latency, and API spend	Advanced
🔄 19	Graceful Degradation	Fallback hierarchy when AI fails: cache → template → error	User-facing system where downtime impacts real people	Advanced
📦 20	Prompt Versioning	Track, compare, and roll back prompt versions like code	Prompts are core product and changes need auditability	Advanced

"My output is wrong" → Start with Chain-of-Thought (#1) + Self-Consistency (#6)

"My output format is inconsistent" → Output Formatting (#4) + Few-Shot Examples (#2)

"My output tone is off" → Role Assignment (#3) + Negative Examples (#8)

"I'm going to production" → System Prompt Architecture (#11) + Adversarial Guardrails (#12) + Evaluation (#16)