The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from construction documents (PDF, TXT, MD) and processes it using an LLM, creating a potential attack surface for indirect prompt injection.
Ingestion points: The DocumentIndexingPipeline class in SKILL.md reads files from the local filesystem and chunks them for indexing into a vector database.
Boundary markers: The ConstructionRAG.answer_query method includes a prompt that instructs the LLM to 'answer based only on the provided context', which is a weak boundary that could be overridden by adversarial content.
Capability inventory: The skill has the ability to read arbitrary files from the filesystem (declared in claw.json) and generate responses using the OpenAI client.
Sanitization: There is no evidence of sanitization or filtering applied to the extracted text before it is used to build context for the LLM.
[EXTERNAL_DOWNLOADS]: The SentenceTransformer initialization in SKILL.md triggers the download of pre-trained embedding models from the Hugging Face model registry, which is a well-known service for machine learning assets.

vector-search