Chain of thought

When to use it

Use it for math, logic, multi-step reasoning, planning or any task where the correct answer depends on getting several intermediate steps right.

When to avoid it

Avoid it for simple factual lookups or single-step tasks — the extra reasoning just adds length and latency without improving accuracy, and can occasionally introduce a wrong turn that wasn't there in a direct answer. With newer reasoning models (models with built-in extended or 'thinking' capabilities), asking explicitly for step-by-step reasoning is often unnecessary — these models already reason internally before answering, and an explicit instruction can sometimes make the response longer without improving its quality. Use plain zero-shot phrasing first with these models, and add explicit chain-of-thought only if the direct answer is unreliable.

Advantages

• Measurably improves accuracy on multi-step reasoning tasks
• Makes the model's logic inspectable — easier to spot where it went wrong
• Combines well with few-shot (showing reasoned examples, not just answers)

Disadvantages

• Longer responses cost more tokens and take longer to generate
• Can occasionally 'reason' its way to a wrong conclusion with confident-sounding steps
• Not useful for tasks that don't actually require multi-step logic