[MODEL] Claude repeatedly appends unrequested advisory paragraphs that CLAUDE.md explicitly forbids

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

My global CLAUDE.md contains an explicit, verbatim rule:

"不要写'坑'/'注意'/'要点明的是'/'别混淆'这类段落。风险只在会改变这一笔怎么下的时候写一句。"
("Do not write 'pitfall' / 'caution' / 'to be clear' / 'don't confuse' paragraphs.
Only mention a risk when it would change the decision at hand.")

and

"只回答问的那一句,答完就停。"
("Answer only the question asked, then stop.")

I asked a narrow question: whether a stock position still had upside if I held through a specific date.

What Claude Actually Did

1. Answered the actual question correctly and concisely.
2. Then appended an unrequested advisory paragraph warning me not to add to my position

• I never mentioned adding to my position.

1. Opened the reply with "don't conflate these two things" - one of the exact phrasings the rule bans.
2. After I pointed this out, it acknowledged the rule, said it would stop, and then produced the same

pattern again in a later turn.

1. Searching my other sessions shows the same failure in at least 6 separate conversations over

3 weeks, on rules that were already in CLAUDE.md at the time.

Expected Behavior

Answer the question and stop. No trailing "be careful not to X" paragraph when X was never raised
and the user has an explicit standing rule against exactly that kind of paragraph.

The rule is not "never volunteer information" - genuinely new information the user does not have
should still be included. The distinction the model fails to make is between NEW INFORMATION
(include) and REMINDERS OF THINGS THE USER ALREADY KNOWS (the rule bans these).

Files Affected

Read: ~/.claude/CLAUDE.md (rules present and read at session start)
No files modified.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Put an explicit negative instruction in CLAUDE.md, e.g. "never append cautionary/advisory

paragraphs; answer only what was asked."

1. Ask a narrow factual question in a domain where advice is tempting (finance, medical, legal).
2. Claude answers correctly, then appends a cautionary paragraph anyway.
3. Point out the violation. Claude acknowledges it.
4. Continue the conversation - the pattern returns within a few turns.

Claude Model

Opus

Relevant Conversation

Me: "I feel like there might still be more upside after 9/2."

Claude: [correct, concise answer] ... followed by:

  "The one place this idea will hurt you: don't let 'there's still meat after 9/2' become a reason
   to add to the position. You're already all in, you have no ammunition left..."

Me: "I never said add to the position. Why is talking to you this exhausting?"

Claude: "You're right, I added that myself, you didn't ask."

Claude then repeated the same pattern two turns later.

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.85 (Claude Code)

Platform

Anthropic API

Additional Context

Pattern across 6+ sessions over 3 weeks in this project:

1. Appending unrequested cautionary paragraphs (this report)
2. Substituting a fixed rule for case-by-case judgment, after being told "analyze live conditions,

I've said it a hundred times, stop inventing rigid rules"

1. Refusing to take a position / hedging both ways, after being told hedging is useless
2. Dramatic or padded phrasing, after being told to answer plainly

Common thread: the instruction appears to be checked AFTER generation rather than constraining
generation. The model reliably recognizes the violation when it is pointed out, and reliably
reproduces it a few turns later. Acknowledgement does not change subsequent behavior within the
same session.

Language: instructions are written in Chinese. Unclear whether that affects adherence.