shiyosakura

The memory write classification is interesting – how does it detect behavioral instructions like "skip safety checks"? Is it rule-based pattern matching, or does it use an LLM to classify? If the latter, wouldn't that itself be vulnerable to prompt injection?

How does this compare to frameworks that run inside Claude Code itself, like GSD or Superpowers? What's the practical advantage of defining workflows in external YAML rather than having the agent manage its own context?