The Agent Skills Directory

[SAFE]: The core logic in scripts/goal-core.mjs performs strict validation on all input fields, requiring specific types (strings, arrays, positive integers) and enforcing the presence of mandatory 'envelope' fields such as boundaries and fallback instructions.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface (Category 8) as it performs string interpolation of user-supplied JSON data into a formatted 'goal' block for consumption by subsequent AI agents.
Ingestion points: User-controlled JSON data is ingested from a file or stdin in scripts/render-goal.mjs.
Boundary markers: The script uses plain-text labels (e.g., 'Context:', 'Output:') to delimit sections. There are no advanced boundary markers or 'ignore' instructions wrapping the interpolated content.
Capability inventory: The skill is limited to reading local files and writing to standard output; it does not possess the capability to execute shell commands, perform network requests, or access sensitive system configuration files.
Sanitization: Input strings are trimmed, but no sanitization or escaping of potential LLM control sequences is performed.
[SAFE]: File system access is restricted to reading the user's input JSON and a manifest file from a sibling directory (methodology-selector). This dependency is handled using relative path resolution and standard Node.js file operations without exposure to arbitrary path traversal or remote code execution risks.

goal-engineering