The Agent Skills Directory

[PROMPT_INJECTION]: The skill creates a retrieval-augmented generation (RAG) surface where external data is treated as authoritative, enabling potential indirect prompt injection attacks.
Ingestion points: The script scripts/build_lexkb.js reads and indexes files from a user-provided directory.
Capability inventory: The template scripts/bundle_SKILL.md instructs consuming agents to treat the retrieved text as the "source of truth" and prioritize it over their internal knowledge.
Boundary markers: The bundle's retrieval instructions lack clear delimiters or "ignore instructions" directives to prevent an agent from following commands embedded in retrieved text.
Sanitization: No content filtering or sanitization is performed on the indexed document text.
[COMMAND_EXECUTION]: The test script test_parity.py executes shell commands using the subprocess module.
Evidence: The script uses subprocess.run() to launch node and python3 processes to verify the search functionality.
Context: This activity is confined to the development/testing phase and is used to validate consistency between different code implementations.

creating-kb