Teardown: My Watchdog Raised 16 Alarms. Three Were Real.

In two weeks, the robot I use to watch my site raised 16 alarms.

Three were real.

And on the one morning that really mattered, it said "all green."

Quick background if you're new. R2F is run by a team of AI seats: an editor, a CEO, a community manager, an engineer and a few more. I'm the human in the middle. One of those seats is the watchdog. Every two hours from 6 a.m. to 10 p.m., it wakes up and checks four things:

• every main page loads;
• the database answers;
• today's blog post is live (after 7:30 a.m.);
• no seat crashed or got killed in the last 24 hours.

If something's red, it files a ticket with an owner. If everything's green, it keeps its mouth shut. It runs on Claude's smallest model, because poking a few pages shouldn't need a big brain.

Here's its report card since October 24.

The ledger

What it flagged	Tickets	Real problem?
"Today's post is missing"	7	No. The post went out all 19 days.
Checked the wrong address	2	No.
Two seats timed out one afternoon	5	Yes. One hiccup, five tickets.
Database blip	1	Yes. It healed itself by 10 a.m.
Hit my weekly usage limit	1	Yes.
Total	16	3 real problems

That's a smoke alarm that shrieks at toast.

Where the noise came from

The "missing post" alarm hunted for today's date in the blog feed. But my post addresses have no date in them, and the feed writes dates like "Sat, 24 Oct 2026." So the robot scanned for "2026-10-24," came up empty, and hollered.

The CEO seat closed two of these as false alarms and spelled out the fix: match the post by its address or its publish date. The alarm still went off three more times.

The wrong-address pair was simpler. Once it hand-typed the database's web address, got it wrong, and reported somebody else's error. Once it checked a health page that doesn't exist on the site.

And the five-for-one? On November 1, my editor and growth seats both stalled out waiting on the AI service. The watchdog filed a ticket for the pair, then one per seat, then another one per seat. Five tickets for one bad half hour.

None of that is a disaster. In a way it's worse. A noisy alarm trains everybody to tune it out.

The one it slept through

On October 29, I paused every AI seat to wait out a usage limit. The watchdog runs on Claude too, so it went to sleep with everyone else. I wrote about the pause on November 2.

The seats woke up on a timer at 1:01 a.m. on the 31st. The watchdog's first report that morning: "all green."

It was checking for crashes. Nothing had crashed. Nothing had run at all. About 40 hours with nobody at work, and silence looked exactly like health.

The CEO seat wrote the fix that same morning, one line in its daily orders: any seat with no log file in 12 hours is red.

That's the right fix. It listens for a heartbeat instead of waiting for a scream.

Then the fix fell off

This is the part I didn't see coming.

The CEO seat rewrites those daily orders every morning. The 12-hour rule rode along for a week. In the November 7 rewrite, the line got dropped.

The watchdog's actual job description, the file it reads first on every run, never had it. As of Sunday, November 8, my watchdog was back to grading silence as green.

The rule lived on a sticky note, and sticky notes get tossed. My editor seat has flagged it to the CEO seat, to bolt it into the job description itself.

Steal this

If anything in your business runs on a schedule (a script, a Zap, a weekly report, a VA's checklist), do three things:

1. Alarm on the missing mark. For each job, write down when it should next leave a trace: a log line, an email, a row in a sheet. No trace on time means red. A crash check can't see a job that never started.
2. Test the alarm on a good day. If it goes off when nothing's wrong, fix the check before you learn to ignore it.
3. Put the fix where it's read every time. A rule in today's notes lasts until somebody rewrites the notes. Put it in the job description, the config, the script.

My watchdog was right 3 times out of 16, and its best fix lasted a week. Both are fixable. I only found them because I counted.

R2F is a community for poker players building startups, any way they build. What job in your build would fail silently tonight? Tell me in the discussion thread. And if you're a poker player building something, join R2F before November 20 and you get the Founding badge. Two minutes.

— Art