I should mention – these posts themselves show what I’m describing. I’ve been speaking observations into voice notes, Claude’s been helping structure them. Not writing for me, but actually working with me.

Co-work

This was the month Claude started doing actual work rather than just helping with work.

Co-work let me give Claude tasks, not requests. Monitor this contract and tell me what changed. Check these files and summarise the updates. Then I could walk away whilst it worked.

It’s not quite background monitoring – you have to set it off on a task, can’t just leave it watching things – but it’s close enough to be useful. Claude would check documents on my hard drive, spot what had changed, give me a proper summary. Sometimes it flagged something irrelevant or missed a detail, but good enough that I stopped checking everything myself.

Suddenly I could hand off things I couldn’t be bothered to watch. Contracts updating, paperwork changing, files being modified. Set Claude on it, come back later, get told what mattered.

This felt properly different. December was about tools understanding context better. January was about tools using that context without me standing over them.

Claude’s presentations

Claude’s presentation outputs became good enough. Not as good as Gamma’s – Gamma’s still better – but good enough that the speed mattered more for internal work.

At the end of a thought process, I could just get Claude to produce a deck. Right there. Gamma would be another ten minutes of switching platforms and setting things up. For external clients, still Gamma. For internal stuff, Claude was fast enough to be worth the quality trade-off.

Wispr broke

This properly annoyed me.

December’s perfect transcription became January’s mess. Wispr stopped writing down what I said and started guessing what I meant. It would “improve” my sentences, smooth out my phrasing, make everything neater.

Except I didn’t ask for neat. I asked for accurate.

I’d say something specific and get back something similar but wrong. Different word, shifted emphasis, my meaning bent sideways. It had gone from transcription to rewriting, somewhere close to what AudioPen does – active improvement rather than capture.

To be fair to Wispr’s customer service, they were very good at diagnosing things. At least at the beginning. Might have been slightly my own fault – audio quality issues – and introducing a Mac tool called Loopback seemed to improve things a bit.

But the tool had learned what polished text looked like and started producing that instead of what I’d actually said. Overfitting maybe, though I’m not sure that’s quite right.

Tools get better, then worse, then better again.

Gemini actually got there

Gemini became properly good. Actually strong, beyond just being useful for Google work.

If you’ve got Google Workspace and haven’t given your team Gemini and told them to use it, what are you doing? It’s good enough and it’s already sitting inside where the work happens.

This forced a decision I’d been putting off. Chat had to be where the AI was. Running Slack made no sense when Teams had Copilot built in and connected properly to everything Microsoft.

I hate Teams. Genuinely hate it. Still do.

But the integration mattered more than my preference for Slack. So we moved. Teams is worse than Slack – objectively worse – but having chat separated from the AI tools had stopped making any sense at all.

Copilot

Still not quite there. Better than December, occasionally useful, never the first choice.

It worked when embedded in Microsoft tools – Teams, Outlook, Word – but struggled as a standalone thing. It needed an application wrapped around it to be any good.

What changed

We’d moved from “do this specific thing” to “watch this and tell me when something happens.” Not all the time – mostly I was still telling the tools exactly what to do – but enough that it felt like a shift.

The team was doing work that wouldn’t have been possible six months ago. Lovable and Replit were still getting pushed hard. People were finding edges I hadn’t found, working out things I hadn’t worked out.

If you’re looking at this as headcount reduction, you’re getting it wrong. The value isn’t cutting people. It’s the people you’ve got doing things they couldn’t do before.

Where things are

Tools are getting less reactive, more proactive. The cycles are real – things improve, plateau, regress. What works now might not work next month.

The ecosystem matters more than the tool. Gemini works because Google. Copilot sort of works because Microsoft. Claude works because I’ve put time into making it work. How it connects to where the work lives matters more than what it can do in isolation.