Step-by-step prompting, scratchpad reasoning, zero-shot CoT.

Chain of Thought — Reasoning Before Answering

Chain-of-Thought (CoT) prompting instructs the LLM to reason step by step before giving an answer. This dramatically improves accuracy on reasoning tasks (math, logic, multi-step problems). Zero-shot CoT: add "Let's think step by step." to the prompt — surprisingly effective. Few-shot CoT: provide examples of reasoning traces. Scratchpad: allow the model to reason in <thinking> tags before final answer. LLMs with extended thinking (Claude 3.7, o1-series) do this internally.

Interactive Pipeline Flow

UnderstandImplementOptimise

❓ Problem — complex question

Complex tasks: multi-step math, logical deduction, code debugging, root cause analysis. Direct answer prompts fail here — the model jumps to conclusions.

Architecture Diagrams

Using scratchpad tags to separate reasoning from answer.

Loading diagram...

Key Design Decisions

⚡"Let's think step by step" is the simplest CoT technique

Zero-shot CoT (appending "Let's think step by step.") improves accuracy on multi-step reasoning by 30-50% on math and logical reasoning benchmarks. It's free — just append to any complex question.

📋Scratchpad separates reasoning from output

For user-facing AI features, you don't want reasoning shown. Use <thinking> tags: the model reasons internally and presents only the final, polished answer. Claude 3.7's extended thinking uses this natively.

🔮Extended thinking models are worth using for hard problems

OpenAI o1/o3 and Claude 3.7 extended thinking allocate extra compute to internal reasoning. For competition math, hard algorithmic problems, and complex code generation, they dramatically outperform standard CoT.

Discussion