For Claude models at least, you can tell to just manually think in the output and it works fine. I do it reguralrly because for creative writing and summarization, they seem to believe they don't need to think at all, and get way worse results.