The fix: ask it to reason step by step

Chain-of-thought prompting tells the model to lay out its reasoning before the final answer. Instead of just the number, it produces a reasoning trace:

Start with 23 apples. They used 20 for pies, leaving 23 minus 20 equals 3. Then they bought 6 more, so 3 plus 6 equals 9. The answer is 9.

Why does this help so much? Because each step the model writes becomes part of what it reads for the next step. By writing "23 minus 20 equals 3," the model puts that intermediate result into its own working memory, and can build on it. The written steps are like scratch paper. The model is literally giving itself more room to compute by thinking out loud.

flowchart TD
    Q[Hard multi-step question] --> Choice{How do we ask?}
    Choice -->|Answer only| Fast[Model jumps to a number<br/>little room to compute] --> Wrong[Often wrong]
    Choice -->|Think step by step| Slow[Model writes each step,<br/>building on the last] --> Right[Much more often correct]