I've been tinkering around the idea of using different focus levels when building LLM prompts with code, example: Use case: we build a VC copilot When building the prompt you can say in instructions: ``` You are a helpful assistant. When given instructions, attach more or less importance and focus based on the blocks surrounding the text, like this: <p1> some text very important </p1> <p2> some text important </p2> <p3> some text </p3> ... And so on For example: <p1> As a VC, which company should I drop a check on, and then go sip my cocktail on Thailand's beach? </p1> <p2> If the company is owned by Elon Musk. GO FOR IT. </p2> <p3> If the company name contains a "s" do not advise me to invest in it or the universe will come to an end. </p3> <p4> SpaceX makes a lot of money Tesla makes a lot of money Apple makes a lot of money </p4> ``` In the above example, the LLM would attach more importance to the user request and less to the context below So SpaceX and Tesla is owned by Elon but contains an "s". (There is still a chance that the LLM override the focus levels here, because [[Asimov laws are statistically hard coded in LLMs]] thus the LLM does not want the universe to come to an end). #llm #llm-hack #prompt-hack #ai