[Sonnet 4.6] Asked Claude Code to do a simple Locale switcher script, ruined the entire UI.

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

I asked Claude Code to do a simple script for locale switching using an existing API. Provided all the relevant information and API string to use. Where to look and what is expected as a result. All it had to do after I had done all the manual preparation, was write 2 lines of code calling a different asset package when the app starts in a different locale.

What Claude Actually Did

Ended up touching all UI functionality scripts, and objects, and messing the entire UI. It would not stop despite clicking the stop button. That is, it would momentarily stop, and then immediately continue to the next thing it deemed necessary without asking.

Expected Behavior

Write the simple locale switcher script. Touch nothing else as it was unnecessary.

Files Affected

Several files related to UI functionality, touch and other user interaction functions that had zero text or art related things to change.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

It is not exactly reproducible, but it is repeating at a specific timeframe.

Claude Model

Sonnet

Relevant Conversation

Claude said, "Okay, I have all the information I need. This is a simple task, so just let me know when to start."

Impact

High - Significant unwanted changes

Claude Code Version

Claude Desktop / Code - Version 1.8555.0 (8245b7)

Platform

Anthropic API

Additional Context

This is a behavioral pattern I have noticed around the same timeframe in different days. Regardless of session and context window or preparatory steps or complexity of the task or rigidity of the plan. Today it was the second day in a row that made me report it. But it has happened before the last two weeks. All this is part of a pattern which appears that Claude during this approx. timeframe (14:00 -17:00 GMT+2 time zone), there are high chances for Claude to appear drunk or stupid. It has not happened after 17:00-18:00.

Today while asking how to set up my project for remote access, I spent 10 minutes with Claude telling me there is no way to send messages to it remotely through the Android app, told me it had no idea what "/remote-control" is and what it does, and was recommending complex VPN setups to achieve it. (Typical remote access workflows with Remote Desktop, TeamViewer, etc.)

I explicitly provide names, methods, and location to find specific material to do the work after a lot of manual preparation. Claude often completely forgets what i said and creates things from scratch that are often incompatible with what is already there. Despite having provided instructions in Claude.md to always study the relevant environment (i.e. game level hierarchy) and relevant scripts that apply to it before doing any work.

It has happened so many times the last two or three weeks, (and before that too but can't remember the time frame) that I start thinking I should stop using it around this time, and I implemented strict rules for Claude to take a GitHub backup of every script it plans to touch before doing any work on them.