TL;DR - For anything with steps or reasoning, add "think step by step" before the answer. This one phrase sharply cuts mistakes.
Why it matters
Without it, the model jumps to a plausible-sounding answer and skips the working - so it errs on math, logic, and multi-part problems. Forcing it to reason out loud catches those errors before the final answer.
Worked example - before / after
Prompt: "A meeting starts at 2:45pm, runs 90 minutes, then a 20-minute break, then 45 minutes. When does it end?"
Cold answer: often wrong - it guesses.
With "think step by step": 2:45 + 90 = 4:15; + 20 = 4:35; + 45 = 5:20pm. Correct, because it showed the steps.
Steal this - when to use it
Use "think step by step" (or "show your reasoning, then answer") for:
- math, dates, logic, comparisons, multi-part questions, decisions with trade-offs.
Skip it for:
- simple lookups or one-line rewrites (it just adds length).Common mistakes (and the fix)
- Using it for trivial tasks. It bloats simple answers. Fix: reserve it for real reasoning.
- Not asking for the steps separately. Fix: "Reason first, then give the final answer on its own line."
- Trusting the reasoning blindly - skim it; it can reason wrongly too.
Good to know
Newer reasoning models - OpenAI's o-series and Claude's extended thinking - do this internally, so you don't always have to ask. But for everyday models (GPT-4-class, Gemini), "think step by step" is still one of the highest-return tricks. Combine it with few-shot: show examples that include the reasoning.