Skip to content
SHRINESHRINE

Chain of Thought

Ratified

Make the model reason before it answers.

The Pattern

Section titled “The Pattern”
  • Get the model to reason through the problem before it concludes
  • Reasoning models (built-in thinking): raise the reasoning-effort setting first (Model Selection); do not prompt for steps
  • Models without built-in reasoning: ask for step-by-step reasoning in the prompt

When to Use

Section titled “When to Use”
  • Complex reasoning tasks
  • Math or logic problems
  • Multi-step decisions
  • When you need to verify the reasoning, not just the answer

When Not to Use

Section titled “When Not to Use”
  • Simple lookups or formatting; overhead without benefit
  • A reasoning model whose effort setting already covers the step
  • Hand-written step lists for a reasoning model; its own plan often beats the prescribed one

Implementation

Section titled “Implementation”

Reasoning models

  • Raise effort for hard steps; lower it for simple ones (Step-Level Routing)
  • Prefer general instructions (“think thoroughly about edge cases”) over prescribed steps (Anthropic)
  • Skip “think step by step”; these models reason internally, and the instruction can hurt (OpenAI)
  • Some current models may decline a prompt that asks them to write their reasoning out in the answer (Anthropic)

Models without built-in reasoning

  • “Think through this step by step”
  • “Before answering, reason through…”
  • “Show your reasoning, then put the final answer in <answer> tags”

Worked Example

Section titled “Worked Example”
  • Illustrative case (hypothetical)
  • Task: decide whether a schema migration is safe to run without downtime
  • Small model, no built-in reasoning: direct prompt answers “safe”; adding “list each table lock the migration takes, then decide” surfaces a full-table lock and flips the answer
  • Reasoning model: the same step list adds nothing; raising effort from low to high and asking it to “consider locking and replication lag” gets the same catch

Why It Works

Section titled “Why It Works”

Models are more accurate when they work through intermediate steps. Writing out steps helps catch errors that would slip through in a single jump to a conclusion. Reasoning models do this internally, so the lever moves from the prompt to the setting.

Anti-patterns

Section titled “Anti-patterns”
  • Using CoT for simple factual lookups (overhead without benefit)
  • Prompting a reasoning model to “think step by step” instead of raising effort
  • Not reading the reasoning (defeats the purpose)
  • Accepting conclusions that don’t follow from the stated reasoning
Section titled “Related”