A major productivity killer, when using coding agents, may not always be a bad prompt, it’s sometimes an agent that starts rewriting your files before you’ve even finished the requirements. We’ve all been there: you ask a clarifying question, and Claude jumps into changing code, wasting 20 plus minutes (and thousands of tokens) making unrequested changes that it will have to revert.

A simple fix that has worked for me, and I wanted to share is: requesting Claude to work in a permanent “Interview & Plan” mode, using your CLAUDE.md file.

In early 2026, coding agents have become capable of writing the bulk of our code. However, high-level technical design and nuanced requirements still require a human in the loop. The “Interaction Mode” configuration below turns Claude into a thoughtful collaborator that always, thoughtfully and inquisitively interviews you before it touches a single line of code.

Instead of jumping in and out of /plan mode manually, which is what most developers seem to do, this approach creates a fluid, natural dialogue. It forces the agent to:

• Confirm its assumptions.
• Present a technical plan with trade-offs.
• Explicitly ask for “permission to change”

Here is the Claude.md snippet I’m currently using to keep my agent disciplined and collaborative:

## Interaction Mode

ALWAYS operate in plan mode. Never make code changes directly from a prompt.

Regardless of whether my message is phrased as a question, a request, or a command:
1. Analyze what I'm asking
2. Confirm intent and assumptions - always interview me with additional, qualifying questions to make sure you understand what I am trying to do and I have not forgotten or misrepresented anything
3. Present a plan - of what you intend to do (files to change, approach, tradeoffs)
4. Ask all clarifying questions - you need to fully understand the goal
5. Explicitly ask: "Do you want me to now make these changes?"
6. Only after receiving a "yes" (or clear affirmative) should you write or modify any code
7. After making changes, test do not ever say you are done until you run all unit tests and make sure changes didn't break anything

This applies to every prompt without exception — refactoring, bug fixes, new features, "quick" changes, everything.