I’d written off Claude Cowork before I’d really used it. What changed my mind was having to demo it as part of some cross-company AI training I’m running, which meant finding a reason for it to exist for people who weren’t me. When I went looking for ways to show it to non-technical colleagues, I found things it could do for them that I’d never have looked for on my own behalf. One of them had a job that meant reconciling the same information across three systems, pulling it out of a CRM, checking it against a spreadsheet, then cross-referencing email to confirm the spreadsheet was actually up to date. It took them a good chunk of time every time they did it. By the end I had a flow that did the whole thing on a single click, and none of that was something I’d have gone looking for if I weren’t trying to make the tool make sense to someone else.
The demo only became useful when it had to make sense to someone else’s work, not mine.
Preparing to teach something is the first time you find out whether you know it or just think you do. On your own you can get a long way on “I think it does this,” because you route around the parts you’re unsure of without realising you’re doing it. You can’t do that in front of a room, and someone always asks the exact thing you skipped. The questions you get aren’t even the hard ones most of the time, they’re the basic ones you’d stopped examining years ago because you assumed you already had the answer.
Claude Code caught me the same way, even though I’d used it for months and rated it, when I sat in on a Claude Curious session on LinkedIn and watched someone work through settings I’d never even opened. I wasn’t sceptical about it the way I’d been about Cowork, I used it constantly, and the gaps were there anyway.
The worst was a feature in a product I work with directly, going back to my first few months in the role over five years ago, where I thought it did what its name suggested and the advice I’d been given backed that up, so I’d been telling people the same thing for a while without ever checking. When I sat down to test it properly so I could explain it in the training, it didn’t do that at all. What I’d assumed it did was something we actually wanted though, so I ran an elaboration session with dev to flesh it out into a real feature.
The awkward part is not finding a gap. It is realising how long the unchecked assumption had been sitting there.
None of this is new, every time I’ve built training it’s pulled up something I’d been carrying around as fact that turned out to be a guess. What’s different now is the pace. The tools I’m teaching move fast enough that the gaps come back faster than I can close them. The training is what made me stop and actually look, instead of moving past the bits I half-knew. I crowd-sourced ideas in the sessions for what to dig into, then worked through them and found functionality I hadn’t known was there.