The Agent Skills Directory

[SAFE]: The skill is explicitly restricted to an advisory role. It provides analysis and recommendations but is instructed never to execute tasks, run commands, or implement code directly. It specifically checks if the user wants execution and redirects them if so.
[DATA_EXFILTRATION]: Phase 0 (Project Detection) involves reading local project metadata files such as package.json, go.mod, and pyproject.toml to identify the technology stack. While this involves file system access, it targets standard non-sensitive project markers for context. No sensitive files like .env or SSH keys are targeted.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided draft prompts for optimization, creating a surface for indirect prompt injection. 1. Ingestion points: User input draft prompt analyzed in Phases 1-4. 2. Boundary markers: Explicit instructions to remain in an advisory role only. 3. Capability inventory: No subprocess calls, file-write, or network operations detected in the skill instructions. 4. Sanitization: None observed. The risk is minimized as the output is advisory text intended for human review.

prompt-optimizer