The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to perform file operations, specifically writing project memory files and editing global/project instructions (CLAUDE.md). These actions are core to its functionality as a context operator and include safety mechanisms like mandatory diff previews.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it processes session transcripts to extract and persist rules. 1. Ingestion points: Reads the current session transcript for trigger phrases (e.g., '从今以后', '记住') via a hook or manual command. 2. Boundary markers: None explicitly defined in the instructions for separating data from instructions. 3. Capability inventory: Ability to write memory files and edit CLAUDE.md files. 4. Sanitization: Implements a safety protocol requiring the agent to show a diff and obtain explicit user confirmation before applying changes to configuration files.
[DATA_EXFILTRATION]: No network exfiltration or unauthorized data access was detected. The skill resolves the agent's home directory from environment variables or a specific configuration file (~/.skill-publisher/skills/profile.json) to determine where to store context data, which is standard behavior for its described purpose.

lov-auto-context