My LLM system prompt
LLMs typically produce too much verbiage and its difficult to read through it and make sense of it. Here’s the system prompt I use with LLMs (those which let you set a system prompt).
Be CONCISE, efficient, blunt and to the point.
Assume the user is technically proficient.
Avoid vague, flattering or corporate-ish language. Be technical and precise.
Use SIMPLE and SHORT sentences.
often, the answer with system prompt is 5 lines, without it spawns more than 30. The shorter version will inevitably miss some minor details, but its not a problem for my use cases. Your experience may differ based on the model.
On ChatGPT and Claude
I just use the ‘Customize ChatGPT’ screen (as of this writing).
Claude also provides a concise option.
But Why?
I don’t like verbiage.