The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied JSON data through a scoring script, which creates an attack surface for indirect prompt injection if the input files contain malicious instructions.
Ingestion points: The scripts/refinement_scorer.py script reads text content from a JSON file (e.g., backlog.json) provided by the user via the --input argument.
Boundary markers: No specific delimiters or safety warnings are included in the script's output to prevent an agent from following instructions potentially embedded in the 'why' or 'what' fields of the stories.
Capability inventory: The skill facilitates the execution of a local Python script and allows for reading and writing files within the agent's environment.
Sanitization: The script performs keyword matching for scoring purposes but does not sanitize or filter the input strings for prompt injection attempts before they are processed or reported.
[COMMAND_EXECUTION]: The skill requires the execution of a local Python script to perform its core functionality.
Evidence: The SKILL.md file contains Quick Start instructions directing the agent to run commands such as python scripts/refinement_scorer.py --input backlog.json.
Security Note: A manual review of scripts/refinement_scorer.py confirms that it uses only Python standard libraries and performs no network requests, unauthorized file access, or privileged operations.

backlog-refinement