When a prompt is not producing the desired output or when the same prompt is to be used on a variety of texts or in a repetitive task, it pays to spend some time refining the prompt. Depending on the type of problem, different approaches can be attempted. ## The model is not following part of the instructions 1. check the use of [[use delimiters|delimiters]] to separate blocks of text from instructions 2. check if the missed instructions are negative, the model [[prefer positive instructions|prefers positive instructions]]. 3. [[use synonyms]] to replace words in missed instructions 4. [[use causative language]] to replace the text of the missed instructions 5. [[use powerful verbs]] in the instructions ## The model is producing inaccurate responses 1. look for negative instruction, [[prefer positive instructions]] 2. check if the desired outcome is clearly specified (do not [[use synonyms]]) 3. check if you need to [[provide more context]] 4. ask the model to [[check conditions]] before performing the task 5. provide instructions in [[use numbered steps]] 6. ask the model to generate intermediate reasoning steps using [[chain of thought prompting]] 7. provide examples of the desired responses with [[few-shot prompting]] ## The model's output needs tunning 1. if the output is too long, ask for a [[summarization]] 2. [[use causative language]] to make the model more creative 3. use [[tone transformation]] 4. use [[role prompts]] to adjust the output to the user See an example of iterative prompt development in the task of [[generating marketing from the product fact sheet]]. [[how to create a great prompt]] < [[Hands-on LLMs]]/[[5 Prompting]] > [[how to avoid excessive words]]