AI Prompting

Few days ago I posted a video about Claude 4.8 and how it had mixed reviews. Well, I've now been using 4.8 for three days and I gotta say, kind of like it.

Someone talked about why I like 4.8 so much and why it might have got mixed reviews at first. When I first started using 4.8, I did feel it was a little verbose, sterile, academic, but over the last three days I've realized it's really smart.

I also think they might have changed the system prompt after release. A lot of you are pointing out that it seems like it was forcing disagreement with you to avoid being sycophantic, and it would give these overly long explanations of why it was doing what it was doing.

This is just anecdotal, of course, but I feel like that stopped after the first day or two. Where 4.8 really shines is when you give it a clear goal.

If you try to use 4.8 the way you used dumber models, where you give it explicit step by step instructions, it's gonna overthink the task and just spend all your tokens...

ks that Anthropic put out. With the system card, 4.8 is four times more likely to catch bugs in code that it wrote.

I think another way of saying all this is that Claude 4.8 is the best model at using Claude. I think we're truly starting to see the recursive self improvement flywheel take off.

And in just a few weeks, Anthropic will be releasing their methodos model. Except it won't be the methodos they already showed us.

It's going to be the methodos that has all the advantages that 4.8 has. And yes, it will definitely be incredibly expensive, but it's also the model that's writing almost all of the code at Anthropic.

The model's improving itself. We're using the model to improve the systems.

We're building on the model. I don't know where I'm going with any of this, but it's all really exciting.