Prompting Techniques

Once you know the basic recipe, a few techniques can lift your results even higher. In this lesson, you will learn four of the most useful. Zero shot, few shot, chain of thought, and self checking.

• Zero-shot and few-shot
• Chain of thought
• Self-checking

Zero-shot versus few-shot

Zero shot means you just ask, with no examples. Classify this review as positive or negative. This works well for simple tasks. Few shot means you show a few examples first. Here are three reviews, with their labels. Now classify this new one. The examples teach the exact format, and the exact style you want. Few shot is especially powerful, when the task is unusual, or the output must be consistent.

Few-shot in action

Here is few shot in action. We show the model two examples. A sentence, and the emoji that matches. Then we give a new sentence. The model follows the pattern, and answers with a fitting emoji. We never explained the rule, we simply showed it. The model spotted the pattern, from the examples.

Chain of thought

Some problems need steps. If you ask a model, a tricky word problem, it may jump to a wrong answer. But if you add, let's think step by step, it writes out its reasoning first, and the answer is much more often correct. Why? Because each written step becomes part of the context, so the model can build on its own reasoning, instead of guessing in one leap.

1. Ask a hard questiona multi-step problem
2. Add: think step by stepinvite the reasoning
3. Model writes its stepseach step becomes context
4. Better final answerfewer jumps to wrong answers

Self-checking and role switching

Two more tricks. First, ask the model to check itself. After it answers, say, review your answer for mistakes, and correct any you find. It often catches its own errors. Second, ask it to play a critic. Write the email, then act as a strict editor, and improve it. Having the model argue with itself, often gives a polished result.

Summary