A routine is the difference between an agent you use and an agent that works while you are asleep. It is also the fastest way to spend money on nothing, so the order matters: get the job right by hand first, then schedule it.
When a job is ready to be scheduled
Three things have to be true, and if any one of them is missing the routine will produce output nobody reads.
- You have run it manually at least three times and the answer was useful each time. A routine multiplies whatever the Bot already does, including doing it badly.
- The output has somewhere to go. A briefing nobody opens is worse than no briefing, because it trains you to ignore the channel it arrives in.
- The input actually changes. Scheduling a Bot against something that moves monthly is how you get thirty identical reports.
Scope it tight
The advice every practical guide lands on is the same: business hours, not always on. An agent that wakes every fifteen minutes to discover nothing has changed is paying to learn what a schedule already knew. Daily at seven is almost always better than hourly, and weekly is better than daily for anything that moves at the pace of a market rather than an inbox.
Write the stop condition into the description as well as the schedule. A routine that has nothing to report should say so in one line rather than filling a template, and a Bot told that explicitly will do it.
The single most common routine failure is a daily digest that manufactures three items because it was asked for three. Tell the Bot that a quiet day is information and that reporting it is correct.
The four worth running every morning
These are the jobs where a schedule genuinely beats asking, because the value is in the output already being there when you sit down.
The daily briefing
Pulse reads overnight mail and today's calendar and returns three things, each with the first move, and one line for everything else. It runs at 7am on weekdays. The reason it works as a routine is that its value is entirely in arriving before you start deciding.
Meeting preparation
Brief writes a one page brief for every meeting on today's calendar. Prepared at seven, it is waiting; asked for at two minutes to three, it is not.
Competitor changes
Radar checks pricing, changelog, homepage and careers pages weekly and reports only what moved, with the old text, the new text and the URL. Weekly is right here: daily produces noise, monthly means finding out from a customer.
Accounts at risk
Risk scores accounts on Monday morning and returns the ones leaving, ranked by revenue at risk with one action each. Churn is a lagging number and the signals are not, which is exactly the shape of job a routine suits.
What not to schedule
Anything that sends. A routine that drafts is safe, a routine that sends on your behalf is a decision you have delegated without reading it. Every Bot we publish that touches an inbox or a pipeline drafts and stops, including Reply and Keep, and that boundary is written into the instructions rather than left to the schedule.
Also avoid scheduling anything whose output you have not yet decided how to act on. If you do not know what you would do differently on a Tuesday because of the report, the report does not need to exist on Tuesdays.
Reviewing a routine
Two weeks after you set one, read the last ten runs in a row. You are looking for one thing: how many produced an action. If the answer is under three, either the schedule is too frequent or the job is not one worth automating, and both are cheap to fix at two weeks and expensive to ignore for a year.
Routines and the rest of the build
A scheduled Bot reporting that a page is underperforming is only useful if changing the page is quick. That is the other half of the licence: design templates for Lovable, Bolt, Replit, Claude Code, ChatGPT Codex and Cursor, ready to paste into whichever one you build in.