The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its metadata enrichment feature in scripts/sync.py. It reads content from untrusted local documents and feeds it into a prompt for the Claude CLI to generate summaries and tags.
Ingestion points: Untrusted file content is read from the user-configured KB_SOURCE_DIR in scripts/sync.py.
Boundary markers: The prompt template uses ---DOCUMENT HEAD--- and ---END--- delimiters to isolate document content from instructions.
Capability inventory: The injection surface is limited to writing metadata into the frontmatter of Markdown files; no high-risk system commands are triggered by the LLM output.
Sanitization: The script removes Markdown code fences from the LLM output before updating the vault files.
[EXTERNAL_DOWNLOADS]: The skill performs network requests and downloads as part of its document processing pipeline.
Interacts with the MinerU API (mineru.net) for cloud-based OCR and document conversion fallback when local tools are insufficient.
Installs Python packages and downloads machine learning models (Whisper) from Hugging Face repositories at runtime upon user consent.
[COMMAND_EXECUTION]: The script uses subprocess calls to coordinate several local utilities and generators.
Invokes pandoc for document format conversion and ffmpeg for audio processing.
Shells out to the claude CLI for document OCR and frontmatter enrichment.
Automatically generates and writes a sync.command shell script to the local filesystem on macOS to provide a clickable sync entry point.

local-vault