My inbox watch: the exact prompt that taps me on the shoulder and ignores the rest
- skill in action
- automation

Your inbox gets checked forty times a day and almost nothing in it needs you. The problem is not the volume, it is the checking: every glance at a newsletter you did not ask about is a context switch. Filters help, but a filter that silently files your car insurance quote in a folder is a different kind of problem. I wanted the opposite of a filter: something that watches my inbox, taps me on the shoulder only for mail that actually matters, and says nothing the rest of the time.
The setup
A scheduled bot in MyClawBots runs every two hours. It queries my Gmail inbox for messages newer than three hours (max 30 per run), keeps a watermark file of message IDs it has already seen, and compares each run against it. The interesting part is not the plumbing. It is the two lists inside the prompt: the five kinds of mail that count as important, and everything that gets ignored on sight.
The exact prompt
This is the prompt running in my recurring job, verbatim except that I have redacted my account details:
Check my personal Gmail for anything genuinely important that arrived
since the last check. Use the Gmail triage tool with query
'in:inbox newer_than:3h', max 30.
Steps:
1. Load the seen-ids watermark file (create it as [] if missing). Only
consider messages not in it.
2. From new messages, flag as important ONLY: direct personal
correspondence from a real person, security alerts (logins, password
changes), bills/payments/bank notices, travel confirmations or
changes, anything with a deadline or action required.
3. IGNORE: newsletters (List-Unsubscribe, news@, newsletter senders),
promos/marketing, social notifications (Facebook/LinkedIn/Instagram
digests), shopping deals, automated "activity" summaries.
4. If there is new important mail: report it in chat with a short
plain-English line per item (sender + what it is + why it matters).
Read the full message only if the snippet is unclear.
5. If nothing important: stay silent in chat (routine check, no
notification). Still update the watermark file with all newly seen
message IDs.
6. Update the watermark file after every run regardless.
If Gmail errors, retry a few times; if it still fails, stay silent
this round.
Steps 5 and 6 are the whole design. Step 5 means no "your inbox is fine" ping nineteen times a day; I hear nothing unless something matters. Step 6 is what makes the silence trustworthy: the watermark moves forward every run, so the bot never re-reads or re-reports.
What it actually caught
Three runs from the logs:
| Date | New mail | Outcome |
|---|---|---|
| Oct 2 | 19 | All promos, newsletters, automated digests. Stayed silent. |
| Sep 23 | 26 of 30 | Flagged 4 routine bill/payment notices. |
| Sep 21 | Quote request follow-up | Surfaced my GEICO agent's reply with a $1,093 quote that beat my current premium. That one alone paid for the bot. |
The GEICO catch is the reason the setup survives: one time-sensitive quote reply, in an inbox full of newsletters, reached me the same morning. The rule also works in reverse. Security alerts, bank notices, and one-liner heads-ups (phishing attempts, battery warnings, unrecognized charges) come through as short plain-English lines, because the prompt says exactly what earns a ping.
What broke, and the honest limit
On Sep 14 the scheduled run failed partway: the worker response came back truncated, and the check had to be rerun by hand. Nothing was missed because the watermark had not moved, but it exposed the real weakness of this design: silence is a feature until the watcher itself fails. A bot that errors out and stays silent this round gives you exactly the same experience as a bot with nothing to report. There is no separate alert that the watcher is down.
The second limit is that "important" is whatever you write down. The bot cannot exercise judgment you did not give it. A mailing-list reply and an investment statement got filed as routine in my runs because my rules say they are. If your definition of important is different, the first draft of your prompt will miss things until you tune the two lists. That is a fifteen-minute edit, not a reason the approach fails.
Copy it today
Paste the prompt above into your MyClawBots agent, add your Gmail connection, and schedule it every 2 hours. Two things to carry over: the watermark file (without it, every run re-reports the same mail) and the silence rule (without it, you have built another notification firehose). Flat pricing while you test: Lite $19, Pro $39, Max $79, Team $59 per seat, with a 3-day trial that needs no credit card.