Prompt engineering crafts prompts to improve a large language model’s performance, especially when small wording changes can alter the result dramatically. section 19.1.2
A small wording change can change the answer
In the Alice in Wonderland riddle, an LLM often answers correctly with familiar values but fails after changing the values of M and N. Adding “Please think step by step” can improve performance on reasoning tasks. section 19.1.2
Use prompt engineering as a practical set of checks:
- Test alternate wording instead of assuming one successful prompt is reliable. section 19.1.2
- Add explicit reasoning instructions when the task involves reasoning. section 19.1.2
- Separate the requested knowledge from extra behavior, such as asking for Python sorting explained in a particular style. section 19.1.2
- Treat prompts as task-specific queries rather than substitutes for special-purpose programming or retraining. section 1.8
The technique improves access to capabilities already represented in a foundation model, but its sensitivity to minor rewordings also exposes limits in apparent understanding. section 1.8 section 19.1.2
