Why
Caps, “CRITICALLY IMPORTANT” and threats are a habit from the days of models that skipped instructions. Newer models follow each instruction closely and read anything emphasized as top priority. The result is overreach: the rule gets applied where it wasn’t meant to and crowds out common sense.
A reason gives the model a goal rather than a rule. From the goal it works out the cases you didn’t list.
What it looks like
| Instead of | Write |
|---|---|
NEVER use sudo |
A person without root runs the commands; mark anything that needs sudo so they can decide |
You MUST answer briefly!!! |
The answer is read on a phone — 10 lines at most, code in a separate block |
DO NOT INVENT flags |
Flags only from the <help> block: the server runs an old version with fewer flags than the current docs |
IMPORTANT: use British English |
The reader is a UK-based admin; use British spelling, keep terms and commands as they are |
When a hard rule is needed
When there are no exceptions at all: “don’t repeat passwords and keys in the answer, not even masked”. A reason helps here too: “…the answer is stored in history and logs” — and the model understands that a real value isn’t needed in example code either.