The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted manuscript content (Markdown) and structured research data (JSON/CSV) which could theoretically contain malicious instructions.
Ingestion points: Multiple entry points exist in scripts/audit_citations.py (Markdown manuscripts), scripts/select_reporting_guidelines.py (JSON study profiles), and scripts/validate_claim_evidence.py (CSV evidence matrices).
Boundary markers: The skill mandates a safety boundary in SKILL.md and uses scripts/validate_review_intake.py as a required gate before processing confidential material. Instructions explicitly warn against embedding manuscript prose in reports.
Capability inventory: The bundled tools are limited to local file system read/write operations using Python's standard library. There are no capabilities for network communication, subprocess execution, or dynamic code evaluation in the provided scripts.
Sanitization: All inputs are strictly validated against schemas in scripts/_common.py. Sanitization includes size limits (4MB), row/cell limits for CSVs, rejection of NUL bytes, and duplicate-key detection for JSON files. Reports are designed to emit only identifiers and counts rather than echoing raw manuscript text.
[COMMAND_EXECUTION]: The skill defines a set of CLI utilities (validate_review_intake.py, audit_citations.py, etc.) for local auditing.
Evidence: Scripts utilize argparse for bounded input and pathlib for secure path resolution.
Safety Measures: scripts/_common.py implements safe_input_path and safe_output_path which resolve paths, verify file existence, and explicitly reject symlinks to prevent path traversal or symlink attacks. File writes are performed atomically with restricted permissions (0o600).

peer-review