Instructions that hold up in real work.

A system prompt sets the role and the tone. Then the assistant does real work, week after week, and you find the gaps. Two documents disagree. Someone asks something sensitive. Information is missing. Instructions that hold up say what happens then.

You don't need to predict every case. You need sensible defaults, and a clear rule for when to pause, ask, or hand the work to a person.

If two documents give different answers, an assistant will usually pick one, or blend the two, and not tell anyone. That's the dangerous part: the answer sounds sure and nobody knows there was a choice.

If sources conflict, flag the conflict instead of deciding which one is right. Quote both and ask me.

Better still, say which source wins: the current approved policy over an older guide. But keep the flag. A conflict is usually a sign that a document needs tidying, and you only find out if the assistant tells you.

For anything that leaves the business, such as an email to a tenant, a letter to a landlord or a payment, set the boundary first and the wording second.

That one boundary does more than any amount of clever wording, because a mistake becomes a draft someone corrects rather than a message someone receives. Course one's permissions lesson covers how to set it up.

"Be careful" is not an instruction. Neither is "use your judgement". The assistant can't act on either, because neither says what to do. A strong guardrail names the situation and the action.

Four words often do more than another page about the happy path: ask, flag, stop, escalate. Say which one applies when.

• Ask: when information is missing or the request is unclear.
• Flag: when sources conflict or something looks wrong.
• Stop: when the next step can't be undone.
• Escalate: when it needs a person's judgement, such as a complaint, a legal question or anything sensitive.

Here is an outline to reuse. It builds on the five parts from lesson two, with the gaps spelled out.

ROLE: who it is and who it serves.
GOAL: the one job.
INPUTS: use only these sources. If they conflict, flag it.
PROCESS: the steps, in order.
RULES: if missing, ask. If sensitive, escalate to [person].
BOUNDARIES: draft only. No [action] without approval.
OUTPUT: the format.
CHECK: before replying, confirm you followed the format and flagged anything uncertain.

Easy requests pass. Test the ones that don't: a question about a fee, a request with a detail missing, two documents that disagree, something nobody has answered before, something it should refuse.

When one fails, change the smallest rule that would fix it, then run the same cases again. That way a fix for one case doesn't quietly break another.