I built an alarm to protect my work. It went off 176 times.

I have an alarm that stops me at the end of a work session if I finished something and it only exists in one place. It was built after a bad afternoon last summer, when I went looking for a finished piece of work and found it sitting on this machine and nowhere else. Eight other projects were in the same state. A dead hard drive that week would have taken all of it.

Broke

It was supposed to go off once per problem. I finally counted: 176 times across fifteen work sessions in six weeks. Two of those sessions took 47 and 46 by themselves. Same alarm, same message, while I was in the middle of the work it was complaining about.

Nothing about this ever looked broken. It never failed, never threw an error, never went quiet when it should have spoken up. It was doing its job too many times, which no report anywhere counts as a problem.

The reason turned out to be small. The part that was supposed to remember “I already told you about this” only remembered for about a minute, so every few minutes it was meeting me for the first time.

Here is the part that actually cost me something. By week three I was reading past it. That alarm is the only thing standing between me and losing a week of work, and I had trained myself to swipe it away, because it was almost always wrong.

You have one of these. The notification you dismiss without reading. The report that flags forty things every morning when two matter. The reminder that fires for jobs that already closed. It feels like a small daily annoyance and it is not, because the day it is right, you will treat it exactly like the 175 times it was not.

Pick the one you ignore most. Either make it mean something or turn it off. The middle option, an alarm nobody trusts and nobody has switched off, is the expensive one.

Shipped

I have posts drafted for me before I look at them, and lately I did not like how they were coming out. Every one opened in the middle of the story.

The slow fix is rewriting each one. The fix I made took a few minutes: the page of instructions the drafting works from now says what an opening has to do. One change, in one place, and everything drafted after that follows it.

Worth remembering when something you have automated keeps producing work you have to fix by hand. You are allowed to change the instructions instead of the output. Same with the people who work for you. If you correct the same thing three weeks running, the problem is the instructions they were given, and correcting the work again will never fix it.

Steal this

Be a customer of your own business this week. From your phone, not your desk.

Fill out your own contact form at seven at night and see what comes back and how fast. Book the first slot on your own booking page. Text the number on your truck. Find your business on your phone the way somebody who has never heard of you would, and try to give yourself money.

Write down where it was slow, where it was confusing, and where nothing happened at all. That list is worth more than any advice I can give you, because it is yours and it is specific.

Most owners have never done it. The ones who do usually find at least one thing that has been quietly costing them work for months.

Worth knowing

Meta put out a new AI model this month called Muse Spark 1.3, and there is an unusual way to buy it. Through OpenRouter’s contributor tier, it runs at less than a tenth of the standard price. The catch is stated on the same page: what you send it and what it sends back may be used to improve Meta’s products.

For work with nothing sensitive in it, sorting or summarizing, that is real money saved. For anything carrying customer names, prices or contract terms, the discount is paid for with your customers’ information. Worth knowing which one you are doing before somebody wires it up for you.