AndrewRoxby

Yes, in the text he explicitly argues for categorization and analogizing being two sides of the same coin and functionally equivalent. For instance - what is an ‘embedding’? In itself its an analogy. Lakoff explored similar ground.

Good feedback, thank you. I’ll add some to illustrate.

I drive Claude heavily and have been for some quite some time. Starting with the Opus generations after 4.6, I noticed degradation in Claude's prose - both in long, more diffuse conversations on Claude.AI and in task-focused work happening on Claude Code. Documents were likewise filled with condensed, badly written Claudish.

After enduring it for awhile, I set about diagnosing the exact tics and problems that were bothering me, then experimenting with loops where I'd add specific diagnoses to Claude.MD, assess changes in prose quality, then continue tinkering. After many, many turns of iteration across Claude Code and Claude.ai and now multiple generations of both Opus and Fable, the spec has solidified and been thoroughly battle-tested.

I think it works, and works well. I like Claude, and I like Anthropic. I want them to do well and I think they haven't been doing well in terms of Claude's ability to communicate well with humans. Based on recent public comments by Boris Cherny, they seem to agree with this, while acknowledging that it remains a problem.

I feel confident that if people add this, they'll get an immediate improvement in prose quality in recent model generations. For me, it lets Claude's intelligence shine through while being much, much more transparent and readable.

Hope it's useful! Andrew