The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests raw source material such as PDFs, web articles, and YouTube transcripts to generate and update wiki pages. This untrusted content could contain adversarial instructions aimed at manipulating the agent's behavior.
Ingestion points: The raw/ directory and its subfolders, which are populated via file imports, WebFetch, and WebSearch operations as detailed in references/acquire-and-process.md.
Boundary markers: The instructions and schema do not define clear delimiters or "ignore instructions" wrappers for the ingested source text.
Capability inventory: The skill has access to Bash (shell execution), Write/Edit (file system modification), and WebFetch (network access).
Sanitization: No explicit sanitization or filtering logic is provided for the raw source content before it is processed by the agent.
[COMMAND_EXECUTION]: The skill automates complex tasks by executing local shell commands and a bundled Python script (scripts/upload_wiki_ontology.py). These commands handle directory scaffolding (mkdir), source inventory (find), version control (git), and synchronization of knowledge graphs.
[EXTERNAL_DOWNLOADS]: The skill uses WebFetch and WebSearch to acquire research materials and recommends that the user download several third-party CLI tools (e.g., marker, pandoc, yt-dlp) and Obsidian plugins to complete the environment setup.
[CREDENTIALS_UNSAFE]: The provided synchronization script (scripts/upload_wiki_ontology.py) contains logic to search for the INFRANODUS_API_KEY by scanning various local configuration files, including ~/.claude.json, ~/.cursor/mcp.json, and .mcp.json. While this is intended to facilitate connectivity with the vendor's service, it involves reading sensitive environment and configuration files.

llm-wiki