Prompt engineering is the practice of formulating and refining inputs to a language model so it performs better on a task. Because a generative model predicts likely continuations rather than interpreting prompts conversationally, small wording changes can redirect its output. section 1.8 section 19.1.2
Turning autocomplete into a recipe
A bare prompt such as “How can I make brownies?” can produce a simulated forum exchange. Extending it with “The following brownie recipe is easy to make in just a few steps. You can start by” makes a recipe continuation more likely, producing concrete cooking instructions. section 16.3.1
Useful prompt-design moves include:
- Specify the kind of output required by supplying an appropriate context or opening. section 16.3.1
- Assign a role, such as giving short, helpful answers to questions. section 16.3.1
- Add constraints, such as topics that generated responses should avoid. section 16.3.1
- Test wording variations rather than assuming that a successful prompt will remain reliable. section 19.1.2
Prompting can access different regions of a model’s learned distribution, but responses remain difficult to predict. The same sensitivity that enables useful steering can also cause reasoning failures after minor rewordings or distractors. section 16.3.1 section 19.1.2
