Working with Claude

I’m really enjoying working with Claude. I’ve signed up for the cheapest tier (around 18 pounds a month) and we have been building a solution together for the last week or so. I’ve been using Claude Code which plugs into Visual Studio Code very nicely.

I explain what I want the code to do and Claude makes it for me. Quite often it asks for clarification, then it asks for permission to do things, and then it does them. It works directly on your Visual Studio Project (and your PC if you let it). It seems a bit less anxious to please than ChatGPT and hardly ever tells me that something I’ve suggested is a good idea, which I like a lot. It also seems to get things right first time too, which is really nice.

But the best thing about it is the way that it is metered. The first time I discovered this was a bit of a shock. Claude was in the middle of doing something and it just stopped. Like the cooker thing on the moon in Wallace and Gromit’s “Grand Day Out” film when the money runs out. Turns out I’d used my five hour allocation in about an hour and a half, and I had to wait for my next five hour slot to arrive.

At the time I was a bit cross about this, but now I like it. It means that I plan my work a lot more. If I run out of time in a slot I spend the rest of the time writing a detailed specification, or testing what has been produced. Or, better yet, doing nothing. If Claude was available all the time I don’t think I’d ever stop.

I thought I was being clever when I started giving Claude something to do just after I got up so that my morning slot would end around lunchtime and I could have another five hours worth on the afternoon. But then I discovered that there is also a weekly allocation, and if you hit the end of that you have wait for the next week to come around before you can do anything, which would be a pain. So now I track that carefully.

There is a really useful tool here you can use to track your Claude usage. It works very well, although I ended up with a blank dashboard until I performed the fix described here.

It turns out that your usage is all abut tokens. Everything you send to Claude ends up as a token. Everything Claude sends back to you is sent as tokens. A word in a conversation ends up as a token that is sent to Claude. If you are having a conversation with Claude and you ask it another question the entire previous conversation is sent to Claude along with your new question.

This means that long chats can end up costing a lot. If you are starting something new, start it in a new conversation, don’t tag it onto the end of an existing one. You can also ask Claude to summarise your chat and then use that summary to start a new conversation, which should cut down on token use too.

If you have three things you want Claude to do, ask for them all at once. This reduces the number of times you will have to tell Claude what you want and thus the number of tokens used. For me this seemed counter intuitive at first. When I was using ChatGPT I found that if I asked it to perform multiple tasks it would forget to do some of them. Claude is much better in this respect. It seems to use the additional tasks as extra context too.

Finally, avoid peak hours, which for the UK seems to mean between 13:00 and 19:00 at the moment.

Claude as changed the way I work even more that ChatGPT did. I hear that ChatGPT now has a code writing tool too, I’m looking forward to having a go with that.