The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses several local directories for storing RAG-related metadata, including server configurations and subscription costs.
Evidence: The YAML frontmatter in SKILL.md defines configPaths such as ~/Clawic/data/finances/, ~/Clawic/data/servers/, and ~/Clawic/profile.yaml.
Context: These paths are used to track infrastructure details and cost metrics associated with the RAG pipeline. As these are vendor-specific paths (under the Clawic directory) managed by the skill's author, they are documented as standard functionality for this tool suite.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external documents (e.g., PDFs, wiki pages, transcripts), which creates a surface for indirect prompt injection attacks.
Ingestion points: Untrusted documents are ingested into the RAG corpus, as described in the 'Ingesting hostile sources' section of SKILL.md and the 'Prompt Injection From Indexed Documents' section of security.md.
Boundary markers: security.md recommends using structural separation (fenced, labeled blocks) and system prompts to distinguish data from instructions.
Capability inventory: The skill performs file read/write operations within the ~/Clawic/data/ directory to manage indices and logs.
Sanitization: The skill documents defensive measures such as ingestion-time scanning for malicious Unicode or instruction-shaped text, and programmatic verification of citations.

RAG