The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the web, including articles, social media posts, and PDFs. This creates an attack surface where malicious instructions could be embedded in the ingested content. The skill includes explicit instructions in its query protocols to treat wiki files as evidence rather than instructions and to ignore any commands embedded in sources.
Evidence Chain:
Ingestion points: references/ingestion.md (WebFetch, WebSearch, X.com, PDFs)
Boundary markers: Uses YAML frontmatter and standard Markdown delimiters; includes "ignore embedded instructions" warnings in references/query-lite.md.
Capability inventory: Shell access via Bash, file operations via Read, Write, Edit, Glob, Grep.
Sanitization: Redaction of session metadata and a deterministic "seal" process for project handoffs are implemented.
[DYNAMIC_EXECUTION]: The skill uses the Bash tool to execute local adapters, the llm-wiki CLI, and Python-based extraction tools (for PDF and Wayback Machine content) within temporary virtual environments.
[EXTERNAL_DOWNLOADS]: The skill fetches content from well-known services and specific repositories. It uses api.fxtwitter.com and api.vxtwitter.com as proxies for social media data. It also references the ask-grok-mcp tool from a repository belonging to the skill's author (github.com/nvk/ask-grok-mcp), which is treated as a vendor resource.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool for project management, file system operations, and invoking external utilities like git, pdftotext, and various Python libraries. These operations are core to its function as a knowledge-base compiler and manager.

wiki-manager