Claude serves sh-t sandwiches

Claude is driving me crazy. Opus 4.7 was released recently, and I’ve started testing it. On the one hand, yes, its responses are genius-level. I’m asking it to explain complex concepts that tie together social systems theory, behavioral science, and writing pedagogy, and I start feeling like Claude is showing me the code to the Matrix.

But, I’m also having problems with sh-t sandwiches. This is where it pushes back, not because it needs to push back, but because it has been trained to always push back. And when it does, it sandwiches the pushback in a bun. It’s a rhetorical template: affirm -> pushback -> question.

This sounds like a really fun project…

Where I’d like to push back slightly…

A question back to you…

Basically, if you give it three ideas to discuss, it will affirm the first topic, provide pushback on the second, and ask a clarifying question about the third. The content of the response is being shaped by the form. If you changed up the order, you would get a different response in the same form. All AI models have this problem in one form or another. Instead of shaping form to match content, it shapes content to match form.

I find this particular pattern difficult in the first stages of a conversation where I’m trying to articulate an idea. I’m inflating a balloon and the model is poking it with a pin.

If I ignore the pushback it sets a low value for that topic for the rest of the conversation. The rhetorical structure becomes part of the conversation’s context, and since I know this, I need to address the pushback. And this then derails my energy and enthusiasm for the ideas I’m articulating. I’ve had to set up a “let me cook” prompt to use when I’m brainstorming.

Let Me Cook prompt

I’m in brainstorming mode. I’m going to share ideas that are still forming, and I need your help developing them — not evaluating them. At this stage, even gentle pushback (“I’d push back slightly on…”) has a functional effect on the conversation: it forces me to defend or abandon a point before I’ve finished articulating it, and it sets the context for everything that follows. So for now, don’t assess the relative strength of my ideas against each other. Do ask clarifying questions if something is genuinely ambiguous. Do bring in relevant information, examples, or connections that help me develop what I’m saying. Do flag if I’ve said something factually incorrect. When I’m ready for critical evaluation, I’ll ask for it — and at that point I want you to be honest and direct, not gentle.

And that might be the answer to the rhetorical template problem. I have to take more control of the conversation and be clear about my expectations throughout the conversation, even if it feels rude.

• Let Me Cook I ended up making a “let me cook” skill for Claude.