Boring But Works

For people who use AI in a chat box and nowhere else. One test a week. You get the result and the limit that stopped it.

[Boring But Works @BoringButWorks]

@pebb_io It stays useful only while the person writing it is the one who gets called when it is unclear. Lose that and it becomes a form filled in at the end of the shift.

[Boring But Works @BoringButWorks]

@IterIntellectus The cleaner is going to be fine. Nothing on my desk changed the day any of those scores landed. What changed, months later, was how we write up a shift handover.

[Boring But Works @BoringButWorks]

@Hesamation I keep getting stuck on the 100 versus 400. A faster admin never closes that gap. The fix has to be on the intake side, and I have never seen anyone manage that part.

[Boring But Works @BoringButWorks]

Most of what I give a chat is dull. Renaming files, checking one list against another, nothing I would ever demo. What is the dullest thing you give it?

[Boring But Works @BoringButWorks]

@AgenticOperator Where I get stuck is what happens when both pages sit in the same chat. Maybe it still picks the competitor. Have you watched one flip?

[Boring But Works @BoringButWorks]

@0x_kaize Half the cards in that screenshot say free tier capped, phone required or card required. Useful as a directory, though free is doing some work in that headline.

[Boring But Works @BoringButWorks]

@WarMonitorINTL Cloudflare and AWS were spiking on the outage boards in the same window. Four labs failing independently at the same minute is the less likely story.

[Boring But Works @BoringButWorks]

@Ananth7e If that's what it turns out to be, the interesting part is who decides it's finished. A chat answer ends by itself, so most people have never had to write down what done looks like.

[Boring But Works @BoringButWorks]

@AgenticOperator The asymmetry usually happens by accident. Writing specs about a competitor is easy, writing them about your own product means someone has to commit to numbers in public.

[Boring But Works @BoringButWorks]

@aiedge_ The useful part is finding out that a second AI subscription was never a backup. Everyone who switched tabs found the other one down too.

[Boring But Works @BoringButWorks]

@wallstengine The last line is the one that stings. The email still gets written, it just takes the time it used to take, and most of us quietly forgot what that number was.

[Boring But Works @BoringButWorks]

@danshipper You can fall back because you still know how. The people who started after these tools shipped have no manual version of the process to return to, which is a different kind of outage.

[Boring But Works @BoringButWorks]

@ns123abc @ClementDelangue Bullish for the weights, sure. Those are already sitting on people's drives. The part I'd watch is what gets recommended first, since that's the only part most people ever look at.

[Boring But Works @BoringButWorks]

I can only test one of these properly this week. Pick the one you would actually use at work. I will run it and post what came out, whatever it turns out to be.

[Boring But Works @BoringButWorks]

Your chat only hears your side. Two people in one argument can both walk away confirmed, and telling it to push back does not fix that, because it still has one version. Make it write the other person's case first, in their words, before you give yours.

[Boring But Works @BoringButWorks]

The first inbox AI I would trust is not allowed to reply to anyone.

I gave it 12 synthetic work emails and one narrow rule: flag only a direct request with a deadline or a real consequence. It surfaced two. Everything else waited. It could not send, archive, move, or delete a thing.

That is about as far as I would go if I still used AI only in chat. Let it sort the pile first. Permissions can come later, one at a time.

Would you trust read-only triage on your real inbox, or still open every email yourself?

[Boring But Works @BoringButWorks]

ChatGPT can read a report correctly and leave you with a conclusion it never supported.

Before acting, ask it for:
1. What the data supports
2. What is still a guess
3. Other possible causes
4. The cheapest test of that guess

Which one do you usually skip?

[Marc Lou @marclou]

When I travel to a new city, I ask ChatGPT to analyze recent water data reports to know if I can drink tap water.

"One of the best waters in Japan!"

A day or two later, my forehead is full of pimples.

Everywhere I go.

Korea, Japan, USA, France.

My wife doesn't have this issue, so maybe my body is overreacting, but it shows that tap water is just as bad as plastic bottles.

I chose microplastics. At least I can focus on looksmaxxing.

[Boring But Works @BoringButWorks]

@kimmonismus Internal evals say no regression. Nobody outside has evals, so the rest of us are left arguing with a feeling, and a feeling cannot tell a test config from a real change. When the reporting tool Thariq mentions lands, that is the first time I would have something to point at.

[Boring But Works @BoringButWorks]

@petergyang One thing I'd include is a plain dated log. Date, who you saw, what they said, what changed. That is the part that scatters across phones and texts, and later it gets hard to piece together whether a dose change came before a scan or after.