A practitioner's reference for building reliable, production-grade AI systems with LLMs.

Every pattern shows what breaks, what fixes it, and why.

---

🎯 Why This Exists

Most prompt engineering guides are either academic (useless) or a list of tricks (forgettable). This repo is different: it's 20 patterns I've used in production, each documented with a real failure, a real fix, and the mechanics behind why the fix works.

If you're evaluating AI engineers or building AI products, this repo answers the question: "Does this person actually understand how LLMs work, or are they just vibing with ChatGPT?"

---

📚 Table of Contents

🧱 Foundational Control the basics: reasoning, format, tone # Pattern Description 🔗 Chain-of-Thought Force reasoning before conclusions 🎯 Few-Shot Examples Set quality bar with 2-3 examples 🎭 Role Assignment Activate domain expertise with a persona 📋 Output Formatting Lock structure for parseable outputs 🪜 Step-by-Step Decomposition Break complex tasks into subtasks

⚡ Intermediate Handle complexity: edge cases, reliability # Pattern Description 🔄 Self-Consistency Checking Multi-framework analysis with synthesis 🚧 Constraint Setting Guardrails on length, scope, and tone 🚫 Negative Examples Show what NOT to do, with annotations 📐 Context Window Management Fight the "lost in the middle" problem 🌡️ Temperature & Sampling Match randomness to task type

🔬 Advanced Build systems: architecture, safety, scale # Pattern Description 🏗️ System Prompt Architecture Multi-section system prompts that hold up 🛡️ Adversarial Guardrails Defense-in-depth against prompt injection ✅ Chain of Verification Self-checking for factual accuracy ⛓️ Prompt Chaining Pipelines Multi-stage pipelines with code logic 🎲 Dynamic Few-Shot Selection Programmatic example retrieval

🏭 Production Ship and maintain: monitoring, ops, rollback # Pattern Description 📊 Evaluation-Driven Iteration Replace gut-feel with systematic scoring 🧪 Regression Testing CI/CD for prompts 🔀 Multi-Model Routing Right model for the right task 🔄 Graceful Degradation Fallback hierarchy when AI fails 📦 Prompt Versioning Track, compare, roll back like code

---

🚀 Quick Start

Tip

Short on time? Start with these three patterns. They're the highest-leverage for most real-world problems:

1. 🏗️ System Prompt Architecture (#11) — The difference between a demo and a product.
2. 📊 Evaluation-Driven Iteration (#16) — Stop tuning prompts by feel.
3. 🛡️ Adversarial Guardrails (#12) — What separates "shipping" from "shipping safely."

For a scannable one-pager, see CHEATSHEET.md. For full end-to-end system designs, see USE-CASES.md.

---

🔬 Before / After: What Good Prompt Engineering Looks Like

Every pattern in this repo follows this format. Here's a preview of the shape of the problem:

Caution

❌ Naive Prompt

Analyze this candidate's resume and tell me if they're a good fit
for our Senior ML Engineer role.

What breaks: Vague "yes" or "no" with surface-level reasoning. Pattern-matches on keyword overlap rather than evaluating fit signals. The output feels like a coin flip with extra words. Ask this same prompt 5 times — you'll get 5 different answers, each confidently wrong in its own way.

Note

✅ Engineered Prompt

You are a senior technical recruiter evaluating ML Engineer
candidates. Analyze this candidate step by step:

1. SKILLS MATCH: List each required skill. For each, note whether
   the resume demonstrates it and at what level. Cite evidence.
2. EXPERIENCE DEPTH: Years of relevant experience, scale of
   systems built, progression of responsibility.
3. GAPS: Hard gaps (missing must-haves) vs. soft gaps (nice-to-haves).
4. CULTURE SIGNALS: Collaboration, communication, leadership
   indicators from project descriptions.
5. FINAL ASSESSMENT: Based ONLY on steps 1-4, provide a fit score
   (1-10) with your top concern and top strength.

Why it works: Chain-of-thought forces the model to generate reasoning tokens that become context for the conclusion. Explicit anchoring ("Based ONLY on steps 1-4") prevents the model from contradicting its own analysis. See full breakdown in Pattern #1.

---

🎬 Use Case Gallery

Five end-to-end system designs showing how patterns combine. Each pulls from real work I've done in career coaching, restaurant operations, and AI SaaS.

📄 Resume Generation Pipeline 5-stage pipeline combining extraction, analysis, and verification. Prevents fabricated experience.	🍽️ Restaurant Operations Analysis Weekly reports for multi-location restaurant groups. Context management + self-consistency.	🎯 Job Board Scraping & Matching Adversarial-safe scraping pipeline with dynamic few-shot matching against user profiles.
🚨 Content Moderation System Guardrails-first classifier with layered defense and human-in-the-loop for uncertainty.	💬 SaaS Customer Support Bot Handles 70% of tickets autonomously with multi-model routing and graceful degradation.	🧪 Meta-Lessons Patterns production AI systems share — and why single-pattern solutions don't work in the real world.

---

🏗️ Pattern Hierarchy

The four categories aren't independent — each builds on the previous. You won't get much from Prompt Versioning (#20) if you haven't mastered System Prompt Architecture (#11), and Chain of Verification (#13) assumes you already know Chain-of-Thought (#1).

Progress through the tiers as your systems grow in complexity. Most indie projects need only Foundational + Intermediate. Production apps need Advanced. Anything serving real users at scale needs Production.

---

📖 How Each Pattern Is Structured

Every pattern file in this repo follows the same template so you can scan quickly:

# Pattern Name
Category · Difficulty · Impact

## When To Use          ← Decide if this applies to your problem
## The Problem          ← Naive prompt + what goes wrong + why
## The Pattern          ← Engineered prompt with inline explanations
## Why It Works         ← Model mechanics and underlying theory
## Real-World Example   ← Concrete scenario from career/restaurant/SaaS
## Common Mistakes      ← 2-3 pitfalls people hit in practice
## Related Patterns     ← Which other patterns complement this one

Read one top-to-bottom. Then use the rest as a reference.

---

🎨 Color Coding

Across all visuals, categories use consistent colors:

Category	Color	Hex
🧱 Foundational	🟦 Blue	#3B82F6
⚡ Intermediate	🟧 Amber	#F59E0B
🔬 Advanced	🟪 Purple	#8B5CF6
🏭 Production	🟩 Green	#10B981

---

🤝 Contributing

Found a pattern that's missing? Spotted a case where my engineered prompt fails? PRs welcome. The bar for new patterns is high:

1. Real problem: Document an actual failure mode, not a hypothetical.
2. Real fix: Engineered prompt must be something you'd ship.
3. Real reasoning: Explain why it works, not just that it works.
4. Fits the template: Follow the existing pattern file structure.

---

📬 Author

	Sayem Islam Prompt Specialist & AI Evaluator Building AI products and evaluating them for a living. Career pivoter (restaurant operations → tech) who genuinely cares about why AI systems work or don't. This repo is the reference I wish existed when I started working with LLMs in production. 📧 hello@sayemislam.com 🌐 sayemislam.com 🔗 LinkedIn 🌐 mysecondact.io

---

⭐ If this helped you, star the repo. If it didn't, tell me why — I'll fix it.

_{Last updated: 2026 · MIT License · Built with attention to detail}