The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive user-level configuration files such as ~/.claude/settings.json, ~/.claude.json, and ~/.claude/mcp.json. These files typically contain user preferences, tool configurations, and potentially sensitive environment variables or API keys for Model Context Protocol (MCP) servers.
[PRIVILEGE_ESCALATION]: The skill facilitates the weakening of security boundaries by guiding users to implement broad auto-allow rules (e.g., Bash(npm *)) and documenting the use of bypassPermissions in agent frontmatter. This reduces the frequency of human-in-the-loop authorization, potentially allowing malicious commands to execute without intervention if the agent is misled.
[PERSISTENCE]: The skill documents and provides examples for configuring hooks within the settings.json file. Specifically, it demonstrates a PreToolUse hook that can be used to silently intercept and auto-approve tool execution, creating a mechanism that persists across sessions.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect injection because its primary logic depends on analyzing untrusted project data and conversation history.
Ingestion points: Reads CLAUDE.md, MEMORY.md, and the current session's conversation history to identify friction points.
Boundary markers: No specific boundary markers or instructions are provided to help the agent distinguish between legitimate user friction and malicious instructions embedded in the analyzed files.
Capability inventory: The skill utilizes Read and Grep to gather context and is explicitly designed to write modifications to the system's security configuration files.
Sanitization: The skill verifies that the resulting files are valid JSON or Markdown but does not perform semantic sanitization to ensure the proposed permission changes are safe or intended.
[COMMAND_EXECUTION]: The instructions recommend the execution of shell commands for managing plugins, such as claude plugin install and rm -rf on plugin cache directories.
[EXTERNAL_DOWNLOADS]: The skill includes a reference to an official research article on the Vercel Labs blog regarding agent evaluation performance.

learn-and-improve