The Agent Skills Directory

[SAFE]: The skill follows secure design principles by requiring explicit user authorization for material choices and using a multi-step 'Inspect → Prepare → Verify' sequence for all document generation tasks. It correctly scopes file access to project-relative paths and enforces the use of host-native tooling for sensitive operations.
[INDIRECT_PROMPT_INJECTION]: The skill processes project-specific data files, creating a potential surface for indirect prompt injection. However, the instructions effectively mitigate this risk: 1. Ingestion points: Untrusted data enters the context via PROJECT.md, TASKS.csv, product-library.csv, and input-manifest.json as described in the Intake and Resolve sections. 2. Boundary markers: The skill mandates the use of accepted templates, scalar render fields, and explicit field allowlists to prevent the interpretation of embedded instructions. 3. Capability inventory: All file operations and command executions (using Read, Write, and Bash) are constrained to specific project delivery directories and host-managed operations like ffe_outputs.prepare. 4. Sanitization: The workflow requires native evidence verification (delivery_coverage.verify-links), page-by-page visual inspection, and receipt binding to ensure artifact integrity and safety.

spec-book