The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/build_evidence_inventory.py helper script executes local git commands (such as rev-list and hash-object) via the subprocess module to inspect the project history for exact file matches. The implementation uses list-based arguments without shell=True, which is a safe practice that prevents shell injection.
[PROMPT_INJECTION]: The skill processes content from external website files and reference documents, which presents a surface for indirect prompt injection. While malicious instructions could be embedded in the audited files, the skill's purpose is to surface this data as evidence for human review rather than executing it.
Ingestion points: Local files read from user-provided paths for the site root and reference directories in scripts/build_evidence_inventory.py.
Boundary markers: The script provides evidence leads in JSON format; the instructions in SKILL.md guide the agent to interpret these findings without explicit delimiters for the raw text content.
Capability inventory: The script performs file reads and executes standard git subcommands via subprocess in scripts/build_evidence_inventory.py.
Sanitization: The script tokenizes and shingles text to identify overlaps but does not filter the content for instructions.

audit-reference-originality