Recently, Anthropic shared an interesting idea for boosting model performance with a “think” tool. What makes this interesting is that it gives the model a chance to “stop and think” in the middle of the conversation, instead of only at the beginning like reasoning models tend to do.
This approach also works well for getting good results with much cheaper “non-reasoning” models. However, the original post left some of the implementation details a bit vague. In this blog post, I’ll break it down in simple, straightforward terms so you know exactly how to put it to work in Swift!
