The Agent Skills Directory

[PROMPT_INJECTION]: The skill's architecture involves processing untrusted external text documents ('materials') to perform its audit functions. This creates a surface for indirect prompt injection, as malicious instructions embedded within the analyzed project files could potentially influence the agent's consistency reports or subsequent document modifications.
Ingestion points: Text is read from external files provided via the --materials argument in the scripts/consistency_audit.py script.
Boundary markers: The system uses a predefined 'db09' fact source (YAML/Markdown) as a reference, providing a semantic boundary for the audit, although it does not isolate analyzed content from the agent's instruction stream.
Capability inventory: The audit script performs read operations on local paths and writes to a JSON output file; the agent's instructions suggest using the output to apply corrections to other documents.
Sanitization: The script uses regular expressions and text parsing to find terms but does not specifically sanitize the input for hidden agent instructions.
[EXTERNAL_DOWNLOADS]: The references.md file contains numerous links to well-known technology services and documentation platforms (such as GitHub, Amazon, W3C, and Prettier) used for research and methodology background. These references are documented neutrally and do not involve runtime script downloads or execution from untrusted sources.

light-consistency