The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted data as it is designed to analyze claims and sources for potential hallucinations.
Ingestion points: The skill instructions in prompts/llm-hallucination-testing.md and SKILL.md direct the agent to ingest 'LLM hallucination material' and 'available evidence'.
Boundary markers: The skill mitigates risks by requiring a structured 'Input Audit' to record assumptions and missing material, and by separating facts, inferences, and human decisions in the output.
Capability inventory: The skill does not include instructions for writing to the file system or performing network operations based on the ingested data.
Sanitization: The skill enforces the use of specific enums (e.g., SUPPORTED, CONTRADICTED, UNASSESSABLE) for claim validation, reducing the likelihood of the agent following instructions embedded in the analyzed data.
[COMMAND_EXECUTION]: The skill package includes an evaluation script evals/fixtures/check-domain-decisions.sh which executes a local Python script check_domain_decisions.py to validate test results.
This execution is limited to the skill's internal evaluation fixtures and is used to verify that the agent's markdown output matches expected data fixtures.
The Python script uses standard libraries and does not perform network operations or access sensitive system paths.

llm-hallucination-testing