The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user-controlled instructions from ~/.claude/CLAUDE.md. It suggests using XML section tags (e.g., , ) as semantic boundaries, though it correctly notes these are not enforced security boundaries. Capability inventory includes shell commands (cp, wc, cat) and potential outbound prompt flow to external models for validation. Sanitization is addressed by explicit instructions to redact secrets before evaluation and to preserve safety-critical context during conversion.
[COMMAND_EXECUTION]: Includes example shell commands for checking file sizes (wc, cat) and creating backups (cp) within the ~/.claude/ directory. These are standard utility commands used to provide token usage estimates and ensure data integrity during the conversion process.
[DATA_EXFILTRATION]: While the skill involves a 'Multi-Model Validation' phase that suggests testing instructions against external LLMs, it includes a clear instruction to redact secret values before evaluation to prevent credential exposure.

ai-md