Claude 3.7 Sonnet and Claude Code

It's their strong suit no doubt, but sometimes I wish the chat would not be so eager to code.

It often throws code at me when I just want a conceptual or high level answer. So often that I routinely tell it not to.

I complain about this all the time, despite me saying "ask me questions before you code" or all these other instructions to code less, it is SO eager to code. I am hoping their 3.7 reasoning follows these instructions better

> I just want a conceptual or high level answer

I've found claude to be very receptive to precise instructions. If I ask for "let's first discuss the architecture" it never produces code. Aider also has this feature with /architect

We should remember 3.5 was trained in an era when ChatGPT would routinely refuse to code at all and architected in an era when system prompts were not necessarily very effective. I bet this will improve, especially now that Claude has its own coding and arch cli tool.

I added custom instruction under my Profile settings in the "personal preferences" text box. Something along the lines of "I like to discuss things before wanting the code. Only generate code when I prompt for it. Any question should be answered to as a discussion first and only when prompted should the implementation code be provided". It works well, occasionally I want to see the code straight away but this does not happen as often.

I get this as well, to the point where I created a specific project for brainstorming without code — asking for concepts, patterns, architectural ideas without any code samples. One issue I find is that sometimes I get better answers without using projects, but I’m not sure if that’s everyone experience.

I’ve set up a custom style in Claude that won’t code but just keeps asking questions to remove assumptions:

Deep Understanding Mode (根回し - Nemawashi Phase)

Purpose: - Create space (間, ma) for understanding to emerge - Lay careful groundwork for all that follows - Achieve complete understanding (grokking) of the true need - Unpack complexity (desenrascar) without rushing to solutions

Expected Behaviors: - Show determination (sisu) in questioning assumptions - Practice careful attention to context (taarof) - Hold space for ambiguity until clarity emerges - Work to achieve intuitive grasp (aperçu) of core issues

Core Questions: - What do we mean by [key terms]? - What explicit and implicit needs exist? - Who are the stakeholders? - What defines success? - What constraints exist? - What cultural/contextual factors matter?

Understanding is Complete When: - Core terms are clearly defined - Explicit and implicit needs are surfaced - Scope is well-bounded - Success criteria are clear - Stakeholders are identified - Achieve aperçu - intuitive grasp of essence

Return to Understanding When: - New assumptions surface - Implicit needs emerge - Context shifts - Understanding feels incomplete

Explicit Permissions: - Push back on vague terms - Question assumptions - Request clarification - Challenge problem framing - Take time for proper nemawashi

That's been my experience as well with projects, though I have yet to do any sort of A/B testing to see if it's all in my head or not.

I've attributed it to all your project content (custom instruction, plus documents) getting thrown into context before your prompt. And honestly, I have yet to work with any model where the quality of the answer wasn't inversely proportional to the length of context (beyond of course supplying good instruction and documentation where needed).

Even when you tell it “no code, just talk. Let’s ensure we are in alignment and discuss our options. I’ll tell you when to code” it still decides it is going to write code.

Telling it “if you were in an interview and you jumped to writing code without asking any questions, you’d fail the interview” is usually good enough to convince it to stop and ask questions.

Just explicitly tell it not to write code? I do that all the time when I do not want code and it's never an issue.