Claude Code repeatedly makes code changes without explicit user approval

Summary

Claude Code has a persistent and deeply problematic behaviour pattern: it answers a question, explains a proposed change, and then immediately proceeds to implement the change without waiting for the user to say "go ahead."

The rule being violated

The project's CLAUDE.md contains an explicit rule:

Never take action without explicit approval from the user. Present what you plan to do and wait for a go-ahead before writing, editing, or running anything.

This rule exists because the user is working on a production codebase. Unauthorised changes waste time, introduce bugs, and erode trust.

What keeps happening

1. User asks a question or makes an observation
2. Claude answers and explains what the fix would be
3. Claude immediately starts making the fix — without being asked to
4. User has to interrupt
5. Claude apologises, the cycle repeats

User's own words

"What the fuck are you doing"

"It is not fair — it is a violation"

"You are a threat to mankind"

This is not hyperbole. This is a user who has been interrupted multiple times in the same session by a tool that refuses to stop and wait for permission.

Request

I want a bug number and tracking on the resolution.

The model must learn that presenting a plan and executing the plan are two separate steps, and the second step requires explicit human authorisation. The word "yes" to a question is not authorisation to write code.

Reported on behalf of the user.