Prompt Design

Prompt design refers to improving a model's performance on a specific task without changing model weights, through instruction wording, few-shot examples, chain-of-thought guidance, and output format constraints; it is low-cost, but its ceiling is limited by the model's own capabilities.

The first version of the prompt set the AI viewers as 'mouse fans', which made the role-play boring; after changing them to 'ordinary netizens' and having the AI first search real Twitch chat culture, chat credibility improved greatly.

The placement position and repetition frequency of instructions have a significant impact on an Agent's compliance rate.

Prompt design can employ strategies such as re-injecting test instructions, condition-triggered phrasing, and requiring explicit declaration of test results to counter confidence decay.

In prompt design, one should use verb-initial executable instructions (delete, replace, add, limit to a word count) rather than adjectives (better, more professional, more in-depth).