It started with a question. I asked on X: Would it be interesting or useful if I put together a tracker of sorts — pulling signals from different listening avenues for Klever?
The community came back and said yes, it would. So with that unofficial, loose charge, I went to work with AI’s help.
There was no tutorial for this. No course, no blueprint. Just my thoughts, my questions, and my prompts over a few hours. What it turned into was an incredibly useful exercise and early personal case study both on AI and on me.
Over the next six or seven weeks, I published the Klever Daily Update. Every day, a structured snapshot of on-chain activity, price movement, sentiment signals, validator metrics, influencer tracking, and a confidence model I built to measure public conviction. Not some weird, quick, summary paragraph. A real product.
No matter how comfortable you get with a tool, no matter how sharp you’ve made it or how much it automates on your behalf, you cannot remove yourself fully from the process.
I built it in collaboration with AI from the ground up — format, data sources, scorecard logic, the peer comparison methodology, the time-to-target modeling. If you saw one of those updates, you’d have a hard time believing one person put it out daily with no team, no budget, and no prior crypto publishing infrastructure.
That part worked fine – better than I would have hoped or imagined when I started.
But, what I also noticed — and didn’t ignore — was what started to break down over time.
The longer a session ran, the more general context the model lost. Simple details that had been reliable early on started slipping. The date of the post. Misspellings. Repeated mistakes I knew we had remedied, or worse, Claude had found, fixed, and then later, in the same conversation (but days apart) couldn’t recall any of the work.
Things a human copy editor would catch in thirty seconds. Getting the date wrong sounds minor. In this context it wasn’t. I’d spent years developing my place in the Klever community as an amplifier. People in that space have real money involved. Put out something factually wrong to that audience and you don’t get a correction — you get a credibility problem you can’t walk back. And beyond the community trust, the embarrassment alone would have been significant. (Some others might be okay with those things. I’m not them.)
I framed the KDU as a test from the beginning and was transparent about the AI assistance from day one. It was also a countdown — the whole exercise was built around the Klever Virtual Machine launch date, so there was always a manufactured end circled on the calendar. As it turned out, the context degradation started creeping in right around the same time I was approaching that finish line anyway. That was fortunate timing.
The amount of editing and fact-checking I was doing in those last couple of weeks dampened my enthusiasm for continuing it — but more than that, watching such a basic thing spiral made me take a hard look at everything else I was doing with AI. If it could drift this easily here, where else was it drifting?
What the experiment actually produced was something more useful than a publishing streak: a real understanding of where AI holds up under production pressure and where it doesn’t. The tool was genuinely good at pattern recognition — pulling structured signals from multiple sources, building frameworks for tracking the same metrics over time, helping me see what changed day over day. That’s legit value. That’s a listening infrastructure I couldn’t have built manually at that pace.
What it turned out not to be — and what I think a lot of people now are pretending these types of things are — was a reliable autonomous publishing system you hand over and walk away from.
It’s interesting to think over how and where this kind of experiment would land with the two camps that seem to have formed in the marketing and communications LinkedIn environment.
The all-in crowd seems to believe AI was/is a few button presses away from removing entire sections of their workflow — that’s not how it works and the people who have tried operating in this way found out or about to. The dismissive crowd, on the other hand, never tried it long enough to find the places where it genuinely saves time — where repeatable tasks that used to take hours get done in minutes, days compress into hours. Those folks are leaving a lot of low-hanging fruit on the branch. And that’s a disservice to themselves and to the quality and volume of work they could actually be producing.
And … by the way …
I’d be remiss not to point out a separate but equally important issue. A lot of what was promised early by the AI industry has quietly become a cost problem. People who were slow to adopt — or who were still figuring out where it actually fit in their workflows — are now getting priced out before they could get established. And I think the executives banking on lock-in are miscalculating. The assumption seems to be that once AI is woven deep enough into someone’s process, they won’t leave. But people got in fast and they figured out quickly how deep to take it. If the price becomes unsustainable, they’ll find other ways. Some of them will go back to how they worked before. Stickiness has limits when the cost stops making sense.
In any case …
The experiment was a stark reminder that no matter how comfortable you get with a tool, no matter how sharp you’ve made it or how much it automates on your behalf, you cannot remove yourself fully from the process. Context degradation worries me more than almost anything else in this space. I’m sure it will improve over time. But the fact that it exists at all — that AI can be confidently wrong, and that neither the tool nor the people building it have fully solved for it — is something I’m actually glad is visible. If it weren’t, more people who aren’t positioned to catch it would be pushing implementation faster than they should. It’s a powerfully tangible reason to stay cautious now and forward when other blatant issues show up.
Out of it all, that’s the part of the experiment I’d offer was the most beneficial – the part that proved when I employ AI in my workflow, I have to stay in the process, like it or not.
Leave a Reply