Prompt Engineering Framework

The Five Step Framework

• Task
  • Persona: Expertise to draw from the LLM (e.g. Programmer, Artist, Writer, Scientist, etc).
  • Format: The format of the output.
• Context: What you need from it; constraints, background information, backstory, etc. More is better for the LLM to understand and model its output to your specification.
• References: Optional, but this can further shape the LLM to adhere to your expectations. Each reference is often called a shot (related: Explanation of LLMs Being Few-shot Learners + How Tools are Provided to LLMs). 2 to 5 usually yields the best results.
• Evaluate: Evaluate the output and determine whether the output of the LLM is what you wanted
• Iterate: Tweak your prompt based on what needs improvement.
  • Revisiting the prompt framework (Task, Context, References, Evaluate, Iterate), adding new things that make a better prompt. (e.g. persona, format, etc)
  • Separate the prompt into smaller instructions.
  • Paraphrasing or switching to a task that is similar but different enough to yield an altered result.
  • Introduce constraints.

==Substance is key.==

For Images

Detail is more important here, asides from the things listed above, you should also add specifiers about the size, color, position, and aesthetic, or anything that may be needed to help shape the output image better.

Instruction Following from AI

Set the precedent on the first message, continually remind it, AI usually won’t follow your rules if you don’t tell it to the first time.

##

#ai #ai/realworld #ai/prompt-engineering