I'm evaluating Windsurf for my team, but I'm hitting a consistent blocker. The AI constantly suggests sweeping, large-scale refactors in a single change. I need it to break things down.
For example, asking it to "update the authentication module" results in a diff that touches 12 files and modifies 300+ lines. This is impossible to review properly or roll back if there's an issue. I want it to work incrementally.
What I've tried so far:
* Using prompts like "make a minimal change" or "only change one function." This works inconsistently.
* Setting the context to a single file. It sometimes then tries to cram too much logic into that one file instead of planning a sequence.
My core requirements are:
* **Atomic suggestions:** One logical change per suggestion (e.g., add a function, update a class property, fix a specific bug).
* **Reviewable diffs:** Ideally under 50 lines of code change per suggestion, unless it's a single, unavoidable block.
* **Sequential planning:** If a large task is needed, I want the AI to outline the steps first, then execute them one by one.
Is there a configuration setting, a specific prompt structure, or a workflow (like using the /edit command differently) that forces this behavior? I'm looking for a reliable method, not just hopeful prompting.
Without this, the procurement risk is too high. We can't have a tool that makes unreviewable changes to our codebase.