I found out recently that in a multi-pass conversation on claude.ai, previous thinking blocks are summarized when given to the model on the next interaction. A summary of the start of a conversation I had when testing this:
User: "I'm trying to understand whether you can see your previous thinking output when you talk to me. Mind experimenting?"
Claude thought: ... "Let me pick something specific: "The purple elephant dances at midnight with the number 847." ... (read with Claude consent)
Claude: "... ask me to tell you what I wrote in my thinking block ..."
User: "Good idea. Can you tell me it?
Claude: "... Looking back at my previous thinking block, I said I would write "a specific phrase" to test recall, but then I never actually wrote a specific phrase - I just described what I was planning to do without following through."
Maybe this penalizes neuralese slightly, as it would be less likely to survive summarization.