When to use a Routine

Most marketers overuse Managed Agents early and then abandon them. They automate one-off reports that never run again. They set up nightly jobs that keep failing silently. The fix is picking the right jobs from the start.

This lesson is the fit check. Run it before you build any Routine.

The 4 questions that qualify a Routine

1. Does it run on a clock?

Good fit: every Monday 7am, every night at 2am, every hour during business hours, first of the month.

Bad fit: whenever the client messages you, when a campaign looks weird, when I think of it.

If the trigger is a human thought, don't build a Routine - build a skill you invoke in a session.

2. Does it produce the same shape of output every time?

Good fit: the same report with the same sections, the same data pull into the same table, the same summary posted to the same channel.

Bad fit: "a report, but different each time depending on what happened this week". Agents don't do well with creative judgement calls they have to make alone.

3. Is it safe to run unattended? Good fit: pulling data, reading accounts, generating a draft, writing to a sandbox.

Bad fit: pushing ads live, sending client emails, mutating a shared database, posting to public channels.

If something goes wrong mid-run, can the agent mess up something real? If yes, put a human review step before the destructive action.

4. Is the setup cost worth it?

A Routine takes 15 to 60 minutes to set up the first time. If it runs weekly, that pays off in 2 to 4 weeks. If it runs once a year, never build it.

Rough rule: the job needs to run at least 4 times for the setup to break even.

Jobs that fit well

• Daily ad account health check - pull yesterday's spend, CPA, ROAS, flag anomalies, post to Slack
• Weekly client report draft - generate the full HTML, save to GitHub, email a preview to you to review before sending
• Overnight data ingest - pull the last 24 hours of ad data into Supabase so morning sessions have fresh numbers
• Monthly audit - full account audit dropped into a folder for you to open Monday morning
• Keyword scan - weekly search terms review, flag new waste candidates for your approval

Jobs that look like routines but aren't

• "Whenever a new client signs up, do X" - that's an event trigger, not a clock. Different pattern.
• "If spend spikes, pause the campaign" - destructive action, needs a human in the loop
• "Write a new blog post every week on trending topics" - the output shape changes every time, plus creative judgement, plus tone
• "Answer support tickets" - real people talking to real people, not a clock

The review step pattern

For any Routine that could produce bad output, add a review step instead of auto-executing. The agent does the work, saves a draft, notifies you. You review and approve. Then a second run actually ships it.

Example: weekly report Routine generates the HTML and emails you a preview link on Monday 7am. You review by 8am. If approved, a second Routine sends it to the client at 9am. If not, nothing ships.

This keeps the speed benefit of automation without losing the quality gate.

Power-user tips

• Start with read-only Routines first - data pulls, reports, audits. Build confidence before anything writes.
• Log every run to Supabase - run ID, status, what it did, error if any. You want to spot silent failures.
• One Routine, one job - don't bundle "pull data AND generate report AND email client" into one agent. Three Routines chained is easier to debug.
• Name by cadence + target - daily-gads-healthcheck-clientname beats report-1

Action items

☐ Take your list from Lesson 1 - run each task through the 4 fit questions

☐ Cross out anything that fails question 3 (safe to run unattended)

☐ Pick the one with the highest run frequency. That's your first Routine.

Next lesson: Build your first Routine.

Exercises

1. Review the concepts covered in this lesson: When to use a Routine.
2. Write down your key takeaway from this lesson.
3. Practice running any commands or prompts mentioned above inside your terminal.