A year ago I started using Claude as a daily work tool. Not experimentally — actually, as a serious part of how I think through problems.
Here’s what I’ve noticed.
The first month is the worst
Everyone goes through the same arc. You’re blown away by what it can do. You use it for everything. It fails at something important and you overcorrect into cynicism. Then you find the actual middle ground where it becomes useful.
The failure modes are consistent: it’s confidently wrong about specifics, it can’t verify its own outputs, and it has a tendency toward verbosity that you have to actively suppress. Once you know this, you work around it.
What genuinely changed
First drafts. I used to dread them. The blank page problem is real and I have the abandoned Google Docs to prove it. Now I’ll often talk through a piece out loud (or in text) with the model before writing anything — just to find the angle, the structure, the thing I actually want to say. The draft I write is still mine. But the time to first sentence has collapsed.
Research triage. Not for finding facts — I don’t trust it for that — but for mapping a space quickly. What are the main perspectives on this question? What am I not thinking about? It’s good at revealing the shape of a debate without getting lost in it.
What didn’t change
The actual thinking. The interesting part — synthesising, taking a position, being willing to be wrong in public — that’s still just you. The tool can’t want things. It can’t care about ideas. It can’t write a sentence that surprises itself.
The parts of the work worth doing are still the parts only humans can do. That’s both reassuring and, depending on your mood, a bit sobering.