The Agent Skills Directory

[PERSISTENCE]: The skill provides explicit templates and instructions for creating a systemd user service (obsidian-wiki-sync.service) to run synchronization processes continuously in the background.
[PRIVILEGE_ESCALATION]: Instructions include the use of sudo loginctl enable-linger $USER, a command that requires elevated privileges to allow user processes to persist after a session ends.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from external sources and synthesize it into a knowledge base, creating a vulnerability surface where malicious content could influence future agent actions.
Ingestion points: Raw sources are collected from URLs (via web_extract), PDFs, and user-provided text files, which are saved in the raw/ directory.
Boundary markers: The skill uses YAML frontmatter and custom provenance markers (e.g., ^[raw/articles/...]) for metadata and tracing, but these do not provide security isolation between ingested content and the agent's instructions.
Capability inventory: The agent has the ability to read and write local files, search the filesystem, and execute Python code for linting tasks.
Sanitization: There is no mention of sanitizing external content or instructions to ignore embedded prompts within the source material before it is processed and incorporated into the wiki.
[EXTERNAL_DOWNLOADS]: The skill recommends the installation of external tools from non-whitelisted sources, specifically the obsidian-headless package via NPM and the llm-wiki-compiler repository on GitHub.
[DYNAMIC_EXECUTION]: The Lint section provides a Python code template for the agent to use with execute_code to programmatically scan and analyze wiki files for orphans and broken links.

llm-wiki