The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents tools for ingesting data from external sources such as web pages, GitHub repositories, and databases, which introduces a surface for indirect prompt injection.
Ingestion points: Data loading is performed via classes like SimpleWebPageReader, GithubRepositoryReader, and DatabaseReader (referenced in SKILL.md and references/data_connectors.md).
Boundary markers: Examples demonstrate the use of delimiters in PromptTemplate to isolate context from instructions (referenced in SKILL.md and references/query_engines.md).
Capability inventory: The framework documentation covers network access, file system reading, and database interactions alongside agentic tool usage (referenced across all files).
Sanitization: While not explicitly detailed in the snippets, the framework facilitates structured data handling to reduce injection risks.
[EXTERNAL_DOWNLOADS]: The skill contains installation instructions for standard Python packages from the official PyPI registry, including llama-index, openai, anthropic, chromadb, and pinecone. These are widely recognized and legitimate development tools.
[CREDENTIALS_UNSAFE]: Code snippets for database and API configuration use standard placeholders such as user:pass and your-key. These serve as educational examples and do not expose real credentials.
[METADATA_POISONING]: A discrepancy exists between the version number in the metadata frontmatter (1.0.0) and the version mentioned in the text (v0.14.7), though this appears to be a documentation error rather than a malicious attempt to deceive.

llamaindex