Chain-of-Thought Reasoning Cheat Sheet

Updated 2026-05-25

Next Topic: Chroma Vector Database Cheat Sheet

Chain-of-Thought (CoT) reasoning is a prompt engineering technique that transforms how large language models solve complex problems by explicitly requesting intermediate reasoning steps before generating final answers. Introduced by Google Research in 2022, CoT dramatically improves LLM performance on multi-step reasoning tasks—often by 30–60+ percentage points—by mimicking human problem-solving: breaking down questions, showing work, and building toward solutions iteratively. Unlike direct-answer prompting, CoT makes the model's reasoning process visible and verifiable, enabling better accuracy on mathematical, logical, and symbolic tasks while providing interpretability for debugging and trust. By 2026, CoT is no longer just a prompt trick—it is built into the inference modes of frontier models like GPT-5 thinking, Claude Opus 4 extended thinking, Gemini 3 Pro, and DeepSeek-R1, shifting the practitioner's job from triggering reasoning to managing reasoning budgets and evaluating trace quality.

What This Cheat Sheet Covers

This topic spans 13 focused tables and 69 indexed concepts. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core CoT Prompting TypesTable 2: Zero-Shot CoT Trigger PhrasesTable 3: Advanced Reasoning FrameworksTable 4: Tool-Augmented ReasoningTable 5: Self-Verification and ValidationTable 6: Multimodal ReasoningTable 7: Reasoning Model ArchitecturesTable 8: Training and Fine-TuningTable 9: Reasoning BenchmarksTable 10: Optimization TechniquesTable 11: Common CoT PatternsTable 12: Reasoning with CodeTable 13: Limitations and Best Practices

Table 1: Core CoT Prompting Types

Few-shot and zero-shot prompting form the core starting point for CoT; understanding which variant to use — and how to trigger reasoning effectively — is the foundation for every advanced technique that follows. Few-shot remains the most reliable for consistent performance, while zero-shot covers cases where example authoring is impractical.

Technique	Example	Description
Few-Shot CoT	Q: Roger has 5 balls. He buys 2 more. How many? A: Roger started with 5 balls. 5 + 2 = 7. The answer is 7. Q: A store had 20 apples...	• Provides 2–5 example reasoning chains in the prompt showing step-by-step solutions • the model infers the pattern and applies it to new questions. Most reliable CoT variant for consistent performance.
Zero-Shot CoT	Q: If a train travels 60 mph for 3 hours, how far does it go? A: Let's think step by step.	• Adds the "Let's think step by step" trigger phrase without providing examples • surprisingly effective across models and tasks despite requiring no demonstrations.
Auto-CoT	System clusters diverse questions → generates CoT demos per cluster using Zero-Shot → constructs Few-Shot prompt	Automatically generates reasoning demonstrations by sampling diverse questions and using Zero-Shot CoT, eliminating manual example creation while achieving Few-Shot-level performance.

Table 1: Core CoT Prompting Types

Technique	Example	Description
Few-Shot CoT	Q: Roger has 5 balls. He buys 2 more. How many? A: Roger started with 5 balls. 5 + 2 = 7. The answer is 7. Q: A store had 20 apples...	• Provides 2–5 example reasoning chains in the prompt showing step-by-step solutions • the model infers the pattern and applies it to new questions. Most reliable CoT variant for consistent performance.
Zero-Shot CoT	Q: If a train travels 60 mph for 3 hours, how far does it go? A: Let's think step by step.	• Adds the "Let's think step by step" trigger phrase without providing examples • surprisingly effective across models and tasks despite requiring no demonstrations.
Auto-CoT	System clusters diverse questions → generates CoT demos per cluster using Zero-Shot → constructs Few-Shot prompt	Automatically generates reasoning demonstrations by sampling diverse questions and using Zero-Shot CoT, eliminating manual example creation while achieving Few-Shot-level performance.