I've done some experimentation with a Sales Email prompt template (Agentforce with GPT 4 Turbo).
One of the records was missing, so the flow used in the prompt template was not returning the event check in/out dates and room type.
I've added an instruction for the AI to write "MISSING EVENT DETAILS" and stop (1st image).
Instead, the result (2nd image) was that it generated the email using those words as a placeholder, simply repeating it for every gap in the data.
🤦♂️
The only way it worked was after I've added the instruction to just write "EVENT DETAILS ARE MISSING" as the last instruction in the prompt (3rd image).
👍
The way instructions are phrased and where they appear in the prompt can influence the AI model’s interpretation.
The theory is that the final instruction about how to handle missing details is probably assigned with highest priority or guide the output strongly enough to supersede earlier ones.
So what do you think of giving the AI a final directive on how to deal with missing data as best practice?
Isn't it analogous to the fault path we use in flows?
