The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted academic literature (RIS, CSV, and PDF files) which could potentially contain malicious instructions intended to sway the screening result.
Ingestion points: Data enters the system via scripts/prepare_records.py and scripts/prepare_fulltext.py which parse user-supplied exports.
Boundary markers: The skill uses explicit Instruction/Data Boundary blocks in SKILL.md and agents/screening_reviewer_agent.md, which instruct the agents to treat external content as data only.
Capability inventory: Reviewer agents (agents/screening_reviewer_agent.md) are restricted to the Read and Grep tools and have no network or file-write capabilities.
Sanitization: The library scripts/srlib.py contains a spreadsheet_text function that escapes formula prefixes (=, +, -, @) in generated reports to prevent CSV injection attacks.
[COMMAND_EXECUTION]: The skill provides a suite of Python scripts for bibliographic processing and a JavaScript-based Workflow for dual-reviewer orchestration.
Workflow Generation: The scripts/build_workflow.py script generates JavaScript files by embedding configuration data as safe JSON strings, minimizing the risk of arbitrary code generation.
Local Execution: All script operations are performed on local files within a user-specified work directory, with no remote downloads or unverifiable dependencies executed during the screening process.

sr-screener