The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes untrusted manuscript files (PDF, TeX, Markdown) through multiple subagents, which could contain instructions designed to influence the review process or tool execution.
Ingestion points: Reads external manuscript files specified via the user input path ($0) using pdftotext or direct file reads.
Boundary markers: The instructions for forked subagents lack explicit delimiters or warnings to ignore instructions embedded within the processed manuscript content.
Capability inventory: The skill uses tools like Bash, Agent, and Write, creating a chain where malicious data could influence tool execution or file system writes.
Sanitization: No sanitization of the manuscript content is described before it is processed by the AI subagents or passed to validation scripts.- [COMMAND_EXECUTION]: The skill performs shell operations to process documents and validate findings generated by AI subagents.
Evidence: Instructs the agent to execute pdftotext -layout for PDF text extraction.
Evidence: Executes python3 scripts/validate-findings.py to validate JSON findings and compute unique identifiers, where finding metadata (file, line, locus) is interpolated into command arguments.

seven-pass-review