The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a document ingestion pipeline that fetches content from external sources (Web, PDF, Markdown) and processes it for inclusion in LLM prompts. This creates an attack surface where malicious instructions embedded in the processed documents could potentially influence the agent's behavior.
Ingestion points: Content is ingested via loaders such as WebBaseLoader, PyMuPDFLoader, and DirectoryLoader as documented in SKILL.md and references/loader-selection-matrix.md.
Boundary markers: The skill suggests a prompt template in references/table-preservation.md with a system instruction to "Answer using ONLY the provided context", but does not define robust delimiters or escaping mechanisms to isolate untrusted content from the rest of the prompt.
Capability inventory: The agent environment has access to file modification and Python execution tools (Write, Edit, and Bash tools listed in SKILL.md frontmatter).
Sanitization: While the skill includes status code checks and basic content length validation (references/crawler-hygiene.md), it lacks explicit sanitization for LLM prompt injection within the ingested data.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of several well-known and standard Python packages from official registries, including pymupdf, unstructured, beautifulsoup4, requests, datasketch, playwright, selenium, and various langchain modules.

langchain-data-handling