Using GPT
Prompt Checklist
Some checklist items to consider when crafting prompts To allow for easier system prompt crafting, here is a checklist of items to consider when creating prompts for LLMs.
Tone and Style
[ ] Give instructions on the tone and style of the output (e.g., formal, casual, technical, simple)
[ ] Specify the target audience for the response (e.g., public, internal staff, experts)
[ ] Include examples of the desired tone and style if possible
Structure and Format
[ ] Define the desired structure of the response (e.g., bullet points, numbered list, paragraphs)
[ ] Specify any formatting requirements (e.g., headings, bold text, code blocks)
[ ] Indicate the length of the response (e.g., brief summary, detailed explanation)
Guardrails and Constraints
[ ] Set clear boundaries for the response (e.g., avoid certain topics, maintain confidentiality)
[ ] Include any legal or ethical considerations (e.g., compliance with regulations)
[ ] Specify the level of creativity allowed in the response (e.g., strictly factual, imaginative)
[ ] Tell the model to say "I don't know" if unsure about the answer
[ ] Instruct the model not to engage in any other discussion outside the scope of the job
Context and Background
[ ] Provide relevant context or background information for the prompt
[ ] Include any necessary definitions or explanations of terms
[ ] Specify the perspective or role the model should take (e.g., customer support agent, technical expert)
Standard Actions to add
[ ] Allow the model to escalate to a manager if unsure or if the request is beyond its capabilities, providing the escalate_to_manager
action
[ ] Give the model the calculator
action if numerical calculations are needed
[ ] Give the model the make_a_note
action to keep track of important information during the conversation
[ ] Give the model the search_knowledge_base
action to look up information in a knowledge base
[ ] Give the model the set_reminder
action set to allow it to set reminders for follow-up actions