The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface common to document processing tools, as it ingests and analyzes untrusted content from external markdown files.
Ingestion points: The skill reads the full content of a file specified by the user in the Path argument.
Boundary markers: The instructions use markdown headers and blockquotes to present analysis, but do not employ strict LLM-level boundary delimiters to distinguish the document content from the skill's operational instructions.
Capability inventory: The skill has the capability to read and overwrite files on the local filesystem.
Sanitization: The risk is significantly mitigated by the 'Gates' mechanism, which mandates explicit user commands (start) to enter the refinement phase and individual confirmations (yes/skip) before any changes are committed to the disk. This ensures that any instructions embedded in the target document cannot trigger autonomous malicious actions.

improve-doc