I'm wondering if there's a way to tone down the agentic behavior in Claude's coding process. Generally, it works well for me, but there are times when it just keeps generating code non-stop, which can be hard to manage. Is there a way to adjust this in the prompts or perhaps something I could add to the claude.md? Has anyone attempted something similar?
4 Answers
Do you use a todo_list.md to manage its tasks?
I find it amusing that you want less of that agentic behavior! But I totally get it. Planning and implementing one task at a time or making a list of things for it to tackle works great for me too!
It's not that I want to eliminate it, just a bit more control would be nice at times!
Yeah, it starts to make more sense when you're running Claude on a live server and need to create scripts for deployment.
You might want to use shift+tab a couple of times to enter the planning mode, refine your plan, and jot it down in an .md file with phases and checkboxes. Then, you can implement the project step by step while keeping auto-accept on. Lastly, make sure to ask it to clarify any points that seem unclear.
I need to try that! I've been adding "No code!" to every prompt before I'm happy with the planning.
One option is to run it with auto-accept turned off. You can also instruct it to implement only one change at a time to keep things manageable.
I don't use auto mode either, but I've noticed it sometimes goes off track without asking for confirmation for a while. It can be frustrating!
That's a good tip! I also like to suggest that it propose implementation ideas before diving in.
Yes! I create architectural documents that it can use to derive an implementation plan, working through phases. The agent makes a todo list from that. However, sometimes I want to step in during its actions to tweak its direction without losing work already done. Just wish there was a way to get more detailed without waiting to the end!