Best practices for writing instructions

How to phrase a system prompt so the agent behaves consistently and predictably.

Principle: show, don't tell

Weak: “answer in a friendly tone”. Strong: give an example of a typical answer.

Principle: structure beats eloquence

Use Markdown: H2/H3 headings inside the instruction, lists, and explicit “Context”, “Style” and “What not to do” sections. Models handle a structured prompt better.

Principle: constraints matter more than permissions

List what NOT to do (examples of forbidden topics and formats)
Set the boundaries: “answer only about our company's products”, “do not give legal advice”
This is what prevents off-target answers

Principle: keep versions

Keep older versions of your instructions in Notion or Google Docs — that way you can roll back if a new version performs worse.