Most AI desks fail the same way: every new chat or routine gets a fresh briefing, and the human re-explains the same non-negotiables for the tenth time. Do not send email. Do not publish. Do not touch money. Quiet overnight. Draft first.
That is wasted attention. A good desk should carry standing rules the way a good hire carries house policy. You teach it once. Every digest, watch, and handoff inherits the same spine.
The tip
Write a short standing-rules block into the desk profile (or the bot’s durable memory), not into every routine prompt. Routines should describe the job. Standing rules describe the limits that never change.
Keep the block brutal and short. Five to ten lines beats a manifesto. If a rule is not enforceable, cut it.
What belongs in standing rules
- Outbound gate: draft only for email, Slack, posts, and customer replies unless you approve the exact text in chat.
- Money gate: never charge, refund, buy, or change billing without an explicit yes.
- Quiet hours: weekdays only, daytime window, silence overnight and weekends unless the watch is truly time-critical.
- Publish gate: WordPress and social stay draft until you say publish for that exact piece.
- Identity: when writing as you, first person as you — never as the bot.
Soft tools and hard tools both benefit. A morning digest that inherits “stay quiet when nothing changed” stops becoming daily noise. A webhook fulfillment routine that inherits “notify, don’t auto-send money emails until confirmed” stays safer under load.
Paste-ready standing rules
Drop this into your desk’s standing instructions, then keep routine prompts focused on the work:
Standing rules (always on): 1) Draft-first for anything that leaves the desk (email, Slack, posts, DMs). Do not send/publish unless I approve that exact text in chat. 2) Never spend money, change billing, refund, or buy without an explicit yes in chat. 3) Scheduled checks default to weekdays, roughly 8am–7pm local. Stay quiet overnight and weekends unless I marked the watch time-critical. 4) If auth fails for a connector, tell me once, pause the broken routine when it repeats, and stop inventing “done” outcomes. 5) Prefer connectors over browser re-logins. Prefer silence over filler when there is nothing actionable. 6) When writing as me, use first person as me — never as the assistant.
How to know it worked
You stop typing the same caveats into every routine. New watches inherit the gates. When something is ambiguous, the desk asks once instead of guessing outbound. And when nothing changed, you get silence — which is the point.
If you are building a fuller ops desk (calendar, inbox, fulfillment, standing digests), the same pattern scales: one rules spine, many thin routines. That is how a desk stays useful without becoming another thing you babysit.