Build a multi-step automation that reports back

Advanced · ~20 min · Prerequisites: a scheduled task + a connector

What you'll do

Design a task that pulls data from one place, processes it, summarizes the result, and notifies you — all in a single automated run on whatever schedule you choose.

Plan the steps

Before creating the task, write out the sequence in plain language. A multi-step automation typically looks like:

1. Pull — fetch data from a source (your KB, a connected app, the web)
2. Analyze — filter, compare, or evaluate the data against a condition
3. Summarize — produce a concise output
4. Notify — deliver the result to your inbox (and optionally mark it urgent)

Example plan: "Every Monday morning: pull all issues labeled 'blocked' from GitHub, check if any have been open for more than 3 days, summarize the list with owners and days open, and send it to me as a high-priority notification."

Write your plan in the chat and ask the assistant to confirm it understands each step before you finalize. Catching ambiguity early saves debugging later.

Give it the tools

For the task to access what it needs, tell the assistant which sources to use:

• Your knowledge base — docs, notes, and indexed content it can search automatically.
• A connector — a connected app like GitHub, Slack, or Google Drive. The assistant uses the connection you authorized; see Connecting apps if you haven't linked one yet.
• The web — ask it to search the public web as one of the steps.

You can combine sources in a single task: "search the web for news about [topic], compare it to what's in our KB, and summarize the delta."

Send the result

By default the output lands as a notification in your Inbox. You can customize the delivery in your prompt:

• "Mark it high-priority if there are any blockers."
• "Also create a note in my 'Weekly ops' folder with the full output."
• "If nothing is overdue, skip the notification entirely." (this turns it into a Heartbeat-style check)

The notification subject line comes from the task name, so pick a clear one: "Weekly blocked-issue report."

Verify a run

After the first scheduled run, open your Inbox to check the output. If something looks off:

• Reply in the thread "the GitHub step returned 0 results — I think it's searching the wrong repo" and the assistant will adjust the task.
• Check the task's run history in Tasks to see the exact output from each step and spot where it went wrong.
• Ask the assistant to run it immediately — "Run the blocked-issue report now" — to test a change without waiting for the next scheduled time.

Iterate on the prompt until the output is exactly what you'd write yourself. Once it's right, leave it running.

Try this next