The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's workflow integrates local Python utilities (tools/count_words.py, tools/scan_fingerprints.py, tools/build_review_packet.py) to automate text analysis tasks. These are executed as standard shell commands to process user-provided documents. The tests/ directory contains smoke tests that use subprocess.run to validate these CLI tools, representing a controlled and legitimate use of command execution for quality assurance.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted text documents. It mitigates injection risks through a structured evidence chain: (1) Ingestion points: Documents are parsed by the agent and specific word-counting/fingerprinting tools. (2) Boundary markers: Multi-model reviews use a synchronized BRIEF.md manifest, and instructions specify 'Never Cut' rules for critical facts. (3) Capability inventory: The skill is restricted to local file operations and script execution. (4) Sanitization: The scanning engine uses mask_nonprose to exclude code blocks from style analysis, and agents are instructed to perform a 'self-audit' to ensure their review feedback does not mirror undesirable patterns from the input.
[SAFE]: No malicious code, data exfiltration, or persistence mechanisms were detected. The skill employs security best practices, such as SHA-256 integrity checks for review materials, the use of system temporary directories for intermediate files, and verification of Git status to protect user files from accidental modification.

yzr-writing-review