Working Rules

What this lesson is about

Your context file said who you are. This one says what your agent does. In three decisions — a name, a single clear job, and the rules it never breaks — you produce a file called working-rules.md and a folder for the agent to live in. It is short, and it is meant to be: the value is in choosing a job narrow enough to be done well.

What you'll be able to do

• I can name one job my agent does, narrow enough that I could tell whether it did it well.
• I can write rules my agent always follows, in language specific enough to check.
• I can explain why those rules belong in the system prompt rather than in a knowledge file.
• I can save working-rules.md into my toolkit alongside my context file.

A little background

The three-times-a-week test is how you pick the job. Something you genuinely do that often is something you will keep using the agent for, and a tool you would actually use beats one that merely sounds impressive. Once you have the job, the rules follow from it.

The placement matters as much as the wording. A knowledge file gets searched when a question happens to call for it; a system prompt is on from the first word of every session. Rules your agent must never break belong in the second place. The check that you have written a real rule rather than a wish: read it back and ask whether you could tell, from a single reply, if it had been followed.