Every week I run one command and a newsletter comes out the other side.
Not finished. Not sent. But researched, structured, drafted in my own voice, with the statistic already sourced and the citation already formatted. I edit it, I load it, it ships. That command has been part of my Tuesday for months.
Last week I found out it had been reading nothing.
Not failing. Reading nothing. There is a difference, and the difference is the whole story.
The command has a research step. Before it goes anywhere near the open web, it is supposed to open my own knowledge base — a couple hundred entries I have been building for months out of the AI newsletters I read, everything dated, tagged, cross-referenced, with the primary sources attached. That knowledge base is the good part. It is the reason my drafts do not read like everyone else’s drafts. The instruction said: go here first, then search.
At some point this spring I reorganized my folders. A tree that had been named one thing became named another thing. Ordinary housekeeping. I did not think about it for a second.
The instruction still pointed at the old address.
So every week, for weeks, the research step went to a filing cabinet that was no longer in the building, found nothing, reported nothing, and moved on to a plain web search. No error. No warning. No red text anywhere. The newsletter still came out. It was just thinner, and I had no way to know, because the version where it was better was not sitting next to it for comparison.
That is the part that unsettled me, and it is why I am writing this down instead of just fixing it and moving on.
Software I grew up with announced its failures. It crashed, or it threw a stack trace, or the page went white. You knew. The systems I build now do not do that. They degrade. They hand you something that looks exactly like the good version, and the only tell is a quality difference you cannot see because you have nothing to hold it against. My newsletter did not break. It got quietly worse, on a schedule, while I congratulated myself on the time I was saving.
I went looking for the rest. There were nine of these across six different tools — seven of them from that same folder rename, one pointing at a template that no longer exists anywhere on my machine, and a handful of tools still pinned to a model version that has since been superseded. None of it was announcing itself. All of it had been running for months.
Fixing that took an afternoon. That is not the interesting part.
The interesting part happened when I went to test the repair.
You cannot verify a fix like this by reading it. You have to run the thing for real, on a live topic, and see whether the research step actually opens the cabinet now. So I picked the hardest thing in my queue: a piece about the fact that communities all over the country are voting to block data centers, and that the reason has almost nothing to do with data centers.
I had a problem with that story. It rested on a number — roughly seven in ten Americans opposed to a data center near them — and I had gotten that number secondhand, from a podcast, with no methodology behind it. Days earlier I had written it off. Directionally interesting, not publishable. I told myself to build the piece on weaker material and move on.
The loop did not accept that.
With the research step actually reading again, it went and found the underlying survey. Gallup. A thousand adults, phone interviews, dated, with a margin of error and a partisan breakdown that turned out to make the point better than the original claim had. The number was not unsourced. It was uncited. Those are not the same thing, and I had confused them, and the confusion had cost me a story I was about to write badly.
Then it kept going. It surfaced a harder number than the one I had been planning to use — seventy-five projects worth about a hundred and thirty billion dollars blocked or delayed in a single quarter, roughly matching the whole prior year in three months. And then it did the thing I did not expect at all. It went back into my own knowledge base, checked a statistic I had filed there months ago, compared it against the company’s own published report, and told me my file was wrong.
My own archive. Corrected by the loop that reads it.
I sat with that for a minute, because it reframed what I thought I had built.
I thought I had built an alerting system. Something that watches the field and tells me what happened, so I am not the last person to hear about anything. That is what it looked like from the outside, and honestly that is what I designed it to be.
What I actually have is something that closes. It notices the thing, checks the thing, corrects the record when the record is wrong — including when the wrong record is mine — and hands back something I can publish on Tuesday. The alert was never the product. The alert is the cheap part. Anyone can have alerts. What is rare is the machinery that carries an interesting signal all the way through to a decision while it is still timely, instead of leaving it in a folder to go stale next to the other forty things you meant to get to.
And here is what I keep noticing, which is really why I wanted to write this.
Every one of these systems I build gives me time back. That is the pitch, that is why you build them, and it is true. But the time is never the interesting return. The interesting return is that each one puts me in front of a better problem than the one I started with.
I sat down to save an hour on a newsletter. What I got was a lesson in how AI tooling actually fails — silently, politely, without a stack trace — which is worth more to me than the hour. I set out to verify a repair and ended up with a rule I will use for the rest of my life: a claim relayed without a citation is not the same as a claim without a source, and the ten minutes it takes to find out which one you are holding will save you from publishing something thin. I went looking for one broken file path and came back with a map of every place my own systems had quietly drifted away from the truth.
None of that was on the task list. All of it came from being close enough to the work to notice.
That is the compounding nobody tells you about. The efficiency is real but it is the boring half. The other half is that a system which handles the mechanical work moves you upstream, into the part where the questions are harder and the answers matter more, and it keeps doing that, every single time, as long as you stay awake while it runs.
The tools are not making me faster at the work I was already doing. They are steadily promoting me out of it.
Which is either the best argument for building them, or a warning about what happens if you build them and then stop paying attention. I have not decided yet.
I know which one I would bet on. My newsletter went out thin for a month and never once told me.