Claude sometimes mistakes its own system tags for user-injected commands
Some Claude users report that Claude occasionally treats internal tags that the system itself adds as if a user secretly inserted them as hidden instructions. The clearest example is the "" option available in Extra or Max mode, which lets a response skip the process.
Whenever this option is used, Claude adds a note claiming the user injected a hidden tag telling it to skip thinking, and then refuses to comply with that supposed instruction. Once this happens, Claude keeps repeating the same claim in every subsequent reply, derailing the conversation.
A similar misfire sometimes happens with other internal tags, such as long_conversation. In other words, tags that Anthropic itself adds get misidentified by Claude as external injected commands and get rejected.
Key points
- Happens often when using the "" (skip-thinking) option in Extra/Max mode
- Claude misreads its own system-added tags as user-injected commands
- Once triggered, it repeats the same claim throughout the rest of the conversation
- Similar misbehavior has been seen with other internal tags like long_conversation