Scheduled checks
Give an AI worker a repeating task and it runs a check on a schedule, finishing quietly when all is clear and coming to you when something needs attention.
A scheduled check is a repeating task given to an AI worker that Projik runs. On each date the worker gets a fresh copy of the task, does the check and reports. When everything is fine, the copy finishes on its own and nobody is disturbed. When something needs a person, it comes to you for review.
It suits anything you’d otherwise have to remember to look at, such as last night’s data loads or this week’s sign-up numbers against a threshold. Checks that read your own data use data sources, which need the Business plan.
Set one up#
- Create a task that describes the check, and assign it to an AI worker that Projik runs, such as Atlas-bot or Scribe.
- Open the task and find Repeat. Choose Make it repeat.
- Choose how often under Repeats, the time under At, and when Each copy is due.
- Choose Save.
Repeating tasks covers the schedule choices in full.
The worker needs Auto-accept invitations on. Each copy arrives as an invitation, and with auto-accept off it waits for a person to accept it.
What happens on each date#
- Projik makes a copy of the task and invites the worker to it.
- The worker accepts and Projik starts a run. The worker is told the task is a recurring check and that its reply must end with a result line.
- The worker posts its findings as a comment on the copy. The last line is either “Result: all clear” or “Result: needs attention”.
Then one of two things happens.
| The reply ends with | What happens to the copy | Who hears |
|---|---|---|
| “Result: all clear” | It’s finished at once, even when the worker has Require human review before Done on. | Nobody. The worker’s accepting, commenting and finishing aren’t notified. |
| Anything else | It moves to In review. | The person who set up the repeat. The copy also appears under Waiting for your review on their My Work. |
“Anything else” includes “Result: needs attention”, a reply with no result line, and a check that couldn’t be done: the month’s AI allowance was used up, or the run failed after its second try. A check that can’t run never passes silently.
The worker’s findings stay on each copy either way, so you can open a finished copy to see what it found.
Write the task so the result is clear#
The worker decides between “all clear” and “needs attention” from what your task says, so tell it where the line is.
- Say exactly what to check. “Count yesterday’s failed loads in the load log” is a check. “Keep an eye on the pipeline” isn’t.
- Say what counts as a problem. Give the threshold: “Needs attention if any load failed, or if fewer than 20 ran.”
- Say what to report. Ask for the figures behind the result, so a quiet copy is still useful when you open it later.
- Keep it to one check. Several unrelated checks in one task make one result line hard to trust. Give each its own repeating task.
- Keep it small. A run can make at most 20 data calls, and each result is cut to 200 rows. Ask for counts and totals, not whole tables.
- Don’t ask it to fix anything. Data source tools are read-only. If the worker couldn’t do the check, it’s told to treat that as needing attention.
For example:
Check last night's data loads.
Using the Warehouse data source, count the loads that started between
00:00 and 06:00 today, and how many of them failed.
Needs attention if any load failed, or if fewer than 20 loads ran.
Report both counts, and list the failed loads by name.
Changing the check#
Copies are made from a snapshot of the task: its title, description, assignee, skills, estimate and priority as they were when the repeat was last saved. After you edit the task, open Repeat, choose Change and Save again, so the copies to come pick up your edit.
Limits#
- Three open tasks. A repeat makes no new copy while three of its tasks are still open, and says so on the Repeat card. The task you set the repeat up on counts while it’s open. Finish the ones waiting for review and the next date makes a copy again.
- Workers Projik runs only. A copy given to a worker on GitHub or one with its own agent is ordinary work for that worker, with no quiet finish.
- Each check is a run. On a built-in model it counts against the monthly AI allowance, so a check every day uses more than one every week.
- Stopping. Choose Stop repeating on the Repeat card of the task or any of its copies.