1. The clearer the task description, the better
(1) The more specific and clear the prompt is, the better.
(2) LetChatGPTTo play a professional role
(3) Use separators to clearly distinguish which are output requirements and which are reference content in the prompt.
(4) Specify the steps required to complete the task
(5) Provide an example of the final output
(6) Specify the required output content length
2. Provide reference materials
(1) Provide reference materials for the model, similar to reading comprehension.
(2) The model is required to directly quote the original text in the reference material.
3. Split complex tasks into steps or subtasks
(1) Use intent classification to identify the instructions most relevant to the user’s query
(2) For long conversations, make an appropriate summary or filter in the next prompt to avoid going off topic
(3) For long documents, summarize and split the document. Construct the document summary in a recursive way.
4. Let GPT think more
(1) Tell the model not to rush to give the correct answer.
(2) Allow the model to hide its thinking or reasoning process.
(3) Ask the model directly: “Did you miss anything in your answer?”
5. Use external tools
(1) Allow the model to use external information sources in the prompt and use Embeddings to achieve efficient search
(2) Use code or call external API to obtain more accurate calculation results
(3) Allow the model to execute function calls
6. Systematic experiments, continuous trial and error, and optimization
(1) Refer to a gold standard answer to evaluate the model output
