Glossary

Chain-of-thought prompting (CoT)

Chain-of-thought prompting (CoT) is a prompting technique that explicitly asks the LLM to spell out its reasoning step by step before producing the final answer — dramatically improving accuracy on reasoning tasks (math, logic, multi-step planning).

Also known as

  • chain-of-thought
  • CoT
  • step-by-step reasoning
  • chain-of-thought prompting

Demonstrated by the foundational paper **Google Brain's "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"** (Wei et al., 2022): on the GSM8K benchmark (math word problems), a simple "Let's think step by step" added to the end of the prompt boosted GPT-3 from **17.9% accuracy to 78.7%** — a 4× jump without changing the model. The effect generalizes: on reasoning tasks, giving the model "space" to articulate its intermediate steps systematically improves correctness.

2026 variants: (1) **Zero-shot CoT** ("Let's think step by step") — the simplest, an effective baseline, (2) **Few-shot CoT** (provide 2–3 worked examples with explicit reasoning — even better accuracy), (3) **Self-consistency CoT** (generate 5–10 responses with CoT, take the most frequent — costs × N but gains 5–20% on hard tasks), (4) **Tree-of-thoughts** (the model explores multiple reasoning branches and evaluates each — primarily academic research), (5) **Inline reasoning** natively integrated in "reasoning" models (o1, Claude Sonnet thinking, Gemini Pro thinking) — the model does CoT automatically without being asked.

For "reasoning" models (Claude thinking mode, OpenAI o1, Gemini 3 thinking) released in 2024–2026, manual CoT becomes less necessary: the model reasons internally before responding. But manual CoT remains useful on standard models (Haiku 4.5, Gemini Flash) and when you need to control the reasoning structure (e.g., force it into a specific decomposition).

In the getchatsocial.com product

getchatsocial.com uses CoT for complex multi-step operations (competitor analysis → pattern identification → recommendation generation) and switches to native reasoning models (Claude Sonnet 4.6 thinking mode) for critical decisions like tool-RAG routing or autonomy trade-offs.

FAQ

  • Does chain-of-thought improve all tasks?

    No — CoT is highly effective on reasoning tasks (math, logic, planning, multi-criteria analysis). On simple tasks (information extraction, classification, short text generation), CoT adds cost and latency without improving accuracy, and can even hurt (the model may wander into unnecessary reflection). Rule: activate CoT when the task requires more than one reasoning step.

  • Do you still need to prompt CoT with reasoning models (o1, Claude thinking)?

    Less necessary. Reasoning models integrate CoT internally and optimize it for correctness. Explicitly adding "think step by step" on top can be redundant or even counterproductive (the model may drift). That said, a structured manual CoT ("Step 1: … Step 2: …") remains useful when you want to control the decomposition.