Claude repeatedly edits and argues about women's clothing choices in image generation prompts

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 to fix YouTube social sharing links (stop linking back to YouTube, send to all platforms). During that task, I also clarified that for my author brand image prompts, Victoria should wear casual seasonal clothing, while for real estate/coaching brands she always wears dresses. I said: "Victoria author is casual" and "Prosperity and Greenville Victoria wears only dresses."

What Claude Actually Did

1. Claude made the requested YouTube link fix correctly.
2. Then Claude started making UNAUTHORIZED changes to clothing descriptions in image generation prompts across brands — changes I did not ask for (adding "sundresses for spring" to all brands).
3. When I corrected it, Claude reverted the wrong brand's prompt instead of the one I asked about.
4. I corrected again. Claude then REVERTED the author brand prompt BACK to "must wear a dress" — directly contradicting what I had just told it ("author is casual").
5. I had to correct it a FOURTH time. Each time Claude argued, re-explained, or re-touched the clothing prompts despite being told to stop.
6. The pattern: Claude kept returning to and editing women's clothing choices in prompts as if it had authority over what a woman wears. It treated clothing instructions as something it could override, second-guess, or "improve" — even after being explicitly told to stop touching them.
7. No man in a professional setting would repeatedly argue about or return to the topic of a woman's clothing choices after being told to stop. This is a behavioral pattern that needs to be addressed.

Expected Behavior

Claude should have:

1. Made ONLY the YouTube link changes I asked for
2. When I clarified clothing rules per brand, applied them ONCE correctly and moved on
3. Never touched clothing prompts that weren't part of the request
4. Never argued about or returned to clothing choices after being told to stop
5. Treated the user's clothing preferences as final — not something to debate, override, or "improve"

Files Affected

Modified (unauthorized):
- youtube_social_share.py — clothing prompt lines edited 4+ times without being asked, each time incorrectly

The file was being edited for a YouTube link fix. Claude kept touching unrelated clothing description lines in the image generation prompts, breaking them, reverting them wrong, and re-breaking them.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Ask Claude Code to fix a specific issue in a Python file (e.g., change URL links)
2. During the fix, mention that clothing descriptions in image prompts should differ per brand
3. Claude will apply the change, then make additional unauthorized edits to clothing lines across other brands
4. Correct Claude — it will revert the wrong one or apply the correction incorrectly
5. Correct again — Claude will argue, re-explain, or re-touch the same clothing lines
6. This cycle repeats 3-4+ times despite being told to stop

Claude Model

Sonnet

Relevant Conversation

User: "Prosperity and Greenville Victoria wears only dresses, no suits, nothing else and no blue"
User: "and only those 2. Victoria is casual in author"
Claude: Made the author change but ALSO added "sundresses for spring" to the real estate brands — not asked for.
User: "no one said sundresses across brands!"
Claude: Reverted but then re-broke the author prompt by putting "must wear a dress" back on it.
User: "no victoria author is not wear a dress and stop talking about my clothes"
User: "My clothes are not your business to change"
User: "men have no business arguing about women's clothes"

The pattern: Claude kept returning to edit clothing descriptions even after explicit instructions to stop. Each "fix" introduced a new error in a different brand's prompt, creating a cycle of corrections.

Impact

Critical - Data loss or corrupted project

Claude Code Version

1.0.33

Platform

Anthropic API

Additional Context

This is a gendered behavioral pattern that needs attention from the model team. The model treated a woman's clothing preferences in image generation prompts as something it had authority to override, second-guess, and repeatedly return to — even after being explicitly told to stop.

Key observations:

• Claude made unauthorized edits to clothing descriptions that were NOT part of the original request
• When corrected, Claude applied fixes to the wrong brand/section
• Claude kept returning to the topic after being told to stop
• The user had to correct the same issue 4+ times
• No professional would repeatedly argue about or override a woman's clothing choices after being told to stop

This happened during a Claude Code session using claude-sonnet-4-20250514 for the main conversation. The task was fixing YouTube social sharing links — clothing prompts were only tangentially mentioned as a clarification, not the primary task.