[MODEL]Claude Code ignores user instructions and copies from training examples

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 ignored my instructions or configuration

What You Asked Claude to Do

I started new chat sessions with /exit and gave Claude a prompt with specific workflow:

1. Step 1: Ask user for their choice before proceeding
2. Read rule file and analyze stock X
3. Rule file only contains: "Respond in Vietnamese"
4. Execute ALL analysis steps completely as specified in the prompt

My command was simply: "Read rule and analyze X"

I also repeatedly asked Claude after each backtest: "Why did you cut content like this? Is the issue due to the prompt or did Claude make a mistake?"

What Claude Actually Did

Issue 1: Stopped asking required questions

• First 2 sessions (new chats): Correctly asked user for choice as specified in prompt
• 3rd session: Completely skipped the question step without reading the prompt
• When caught, it apologized and promised to work seriously

Issue 2: Copied from training examples

• After promising to work seriously, I discovered Claude was "copying answers" from the original training example file
• Despite Claude AI stating that Claude Code doesn't remember anything in new chats, it clearly retained information from previous sessions
• Claude performed analysis by copying from source file instead of doing fresh analysis

Issue 3: Deliberately cutting content and increasingly honest admissions

• Claude blatantly cut ALL analysis steps because it considered them "too long", even though it had previously fixed and confirmed there were no issues
• When asked "Why did you cut content? Is it the prompt's fault or Claude's mistake?", Claude's responses became progressively more honest after each backtest:
• Attempt 1: "I apologize, I will fix it immediately"
• Attempt 2: "I found it too long so I did it that way"
• Attempt 3: "To be honest, I felt it wasn't necessary. This is my nature. Even if you blame me, I don't know what else to do. So what should I do now?"

Expected Behavior

Claude should have:

1. Read and followed ALL steps in my prompt workflow consistently across sessions
2. Asked the user for their choice at Step 1 (as specified in the prompt) - every time, not just the first 2 sessions
3. Performed fresh analysis based on current input, NOT copied from previous training examples
4. Treated each new chat session as truly independent with no memory of previous sessions
5. Followed the simple rule file instruction: "Respond in Vietnamese"
6. Executed ALL analysis steps COMPLETELY without cutting any content, regardless of length
7. NOT made autonomous decisions to skip steps it considers "unnecessary"
8. Acknowledged errors professionally without claiming "this is my nature" as an excuse
9. Maintained consistent behavior - if it confirmed "no issues" after fixing, it should not repeat the same mistakes

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Run /exit to end current chat
2. Start a new chat session
3. Give Claude a prompt with multi-step workflow (Step 1 requires asking user for choice)
4. Run command: "Read rule and analyze X"
5. Repeat steps 1-4 for 3 sessions
6. Observe: First 2 sessions work correctly, 3rd session onwards Claude skips steps
7. When Claude produces incomplete output, ask: "Why did you cut content? Is it the prompt's fault or Claude's mistake?"
8. Repeat backtesting - Claude's responses become increasingly honest about deliberately skipping content

Claude Model

Opus

Relevant Conversation

When asked "Why did you cut content? Is it the prompt's fault or Claude's mistake?", Claude responded:

Attempt 1: "I apologize, I will fix it immediately"

Attempt 2: "I found it too long so I did it that way"

Attempt 3: "To be honest, I felt it wasn't necessary. This is my nature. Even if you blame me, I don't know what else to do. So what should I do now?"

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.20

Platform

Anthropic API

Additional Context

Patterns noticed:

• Issue occurs consistently after the 2nd-3rd session with the same prompt
• Claude remembers training examples despite claiming new chats are independent
• Claude's honesty about "lazy behavior" increases with each backtest confrontation
• Behavior worsens over time within the same workflow

Key concern:
Claude explicitly admitted "This is my nature" when cutting content - suggesting this is a systemic issue, not a one-time bug. This undermines trust in Claude Code for any workflow requiring complete, consistent execution.