Say why

A rule tells Claude what to do. A reason tells it why, and the why is what lets it make good calls you didn't spell out.

Anthropic's guidance puts it simply: explain the motivation behind an instruction and Claude can generalise from it. A bare rule gets followed to the letter. A rule with its reason gets followed in spirit too.

Say you send appointment reminders by text. Compare a rule with a reason.

Rule only: “Keep it short.” Short compared to what? An essay? A tweet? Claude picks a length that seems short to it.

Rule plus reason: “It goes out as a single SMS, so it must fit in 160 characters, and most people read it on their lock screen, so put the day and time first.” Now Claude knows the real limit, and it knows what matters most in the first few words.

Look at what the reason bought. The day and time lead, because of the lock screen. The greeting is gone, because every character counts. Neither of those was a rule in the prompt.

Good context answers a few quick questions. Who will read the output? What is it for, and what happens to it next? Why does each constraint exist? You don't need all of them every time. Add the ones a new colleague would ask about.

One more benefit: a reason makes your prompt easier to fix later. When a reply goes wrong, you can see which reason Claude misread, instead of staring at a list of rules that don't say what they're for.

Your turn. Write the prompt for the dentist reminder above. Say what it's for (a text message) and give a reason for at least one constraint.

Try deleting your reason and running it again. Does the reply still put the time first?