Back to DocsPrompt Engineering

Chain-of-Thought Prompts

Force the model to reason step-by-step before committing to an answer. This recipe works across every frontier model and dramatically improves accuracy on multi-hop logic, math, and structured analysis.

The Template

Think step-by-step before answering.

1. Restate the problem in your own words.
2. List every fact or constraint you can identify.
3. Walk through the reasoning chain aloud.
4. Flag any assumptions or missing information.
5. Produce the final answer only after step 4.

Problem: {INSERT}

Why It Works

  • •Prevents premature token commitment — the model cannot blurt an answer before reasoning.
  • •Surfaces hidden assumptions that would otherwise corrupt the output.
  • •Turns a single forward pass into an implicit scratchpad, improving multi-step accuracy by 20-40% in benchmarks.

When to Use

Math word problems, legal clause analysis, debugging traces, multi-constraint planning, and any prompt where the model tends to skip intermediate steps. Skip it for simple factual lookups — the extra tokens aren't worth the cost.

← Zero-Shot ClassificationStructured Output →