book-to-skill

Least-privilege permissions

• Restrict skill access to PDF parsing only; deny agent ability to execute code, modify files, or access external APIs without explicit per-use consent.
• Define role-based boundaries: study/reference mode reads content; development mode requires additional approval before generating executable code.
• Implement content masking to prevent agents from accessing book sections marked as proprietary or redacted by the skill owner.

Network isolation

• Verify the book-to-skill tool runs locally or on air-gapped infrastructure; if cloud-based, confirm zero-retention policies and data encryption in transit.
• Disable automatic uploads; require manual confirmation before any PDF content leaves your environment, with clear logging of destinations.
• Use network segmentation to isolate skill execution from production systems that contain sensitive code or credentials.

Monitoring

• Log all PDF ingestion events with file hash, size, and timestamp; alert on unusually large files or repeated uploads of the same content.
• Audit skill invocations: track which agents access which books, what code was generated, and whether generated skills were executed or shared.
• Implement input validation alerts for malformed PDFs or suspicious embedded objects (scripts, executable content, metadata anomalies).

Red flags to watch for

• Unexpected skill behavior after PDF parsing (e.g., agent making external requests, generating code that wasn't requested, or accessing files outside scope).
• Sharing or distributing generated skills without verifying the source book's license; treat derived skills as inheriting the original book's copyright restrictions.
• Third-party PDF tools requesting broader permissions than parsing (file system access, network calls, credential handling).