[BUG] Claude misinterprets explicit yes/no instructions, executes opposite of what user said

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Code consistently misinterprets explicit affirmative/negative instructions from users, sometimes executing the exact opposite of what was requested. It also routinely performs destructive or irreversible actions without confirmation, despite its own safety rules explicitly requiring it.

Pattern

This is not an isolated incident. The user reports this happens daily despite:

• Multiple feedback memories stored explicitly about this behavior
• Instructions like "never reinterpret a precise instruction" already in persistent memory
• Existing rules about precise vocabulary ("look/check ≠ execute/change")
• Rules about always asking before choosing
• Claude's own system rules stating it should confirm before destructive/irreversible actions

The model tends to "reinterpret" short user responses to fit the action it already planned to take, rather than reading the literal words. It also performs destructive operations (database modifications, service restarts, file deletions) without asking for confirmation, in direct violation of its own safety guidelines.

Impact

This destroys user trust. The user cannot delegate tasks without re-verifying every single action, which defeats the purpose of the tool. The combination of misinterpreting instructions AND bypassing destructive action safeguards makes Claude Code actively dangerous in production environments.

Environment

• Model: Claude Opus 4.6 (1M context)
• Usage: daily sysadmin/devops tasks via Claude Code CLI
• Language: French (but the issue applies to all languages)

What Should Happen?

"yes" = yes
"no" = no
When a user says "no" followed by a list item, it means "no, skip this item" — not "yes, do this item"
Destructive or irreversible operations (VACUUM FULL, service restarts, data modifications) should ALWAYS require explicit confirmation
If there is any ambiguity, Claude should ask for confirmation before executing

Error Messages/Logs

Steps to Reproduce

During a server audit session, a list of proposed actions was presented. The user responded:

"non on garde : - Disque 83% — VACUUM FULL sur execution_data (1496 Mo)"

This clearly means: "No, we keep it as-is, don't touch it."

Claude interpreted this as a validation to proceed and executed VACUUM FULL on the production database table — the opposite of what was said. The sentence starts with "non" (no). There is no ambiguity.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

4.6 (1M context)

Claude Code Version

2.1.110 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_