The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a mechanism where the agent ingests data from local markdown files and user-provided import paths to dynamically update its behavioral rules. 1. Ingestion points: The skill reads from 'MEMORY.md', '.claude/instincts.md', and any file path provided during the 'import' operation. 2. Boundary markers: The system relies on markdown formatting (headers and list items) but lacks explicit instructions to treat file contents as untrusted or to ignore embedded control sequences. 3. Capability inventory: The skill modifies agent behavior globally by instructing it to 'apply rules proactively', and it possesses file write capabilities to update these instruction stores. 4. Sanitization: There is no evidence of input validation or sanitization for imported data, only a numerical confidence decay which does not prevent instruction injection.

instinct-system