walkthrough
Pass
Audited by Gen Agent Trust Hub on Aug 15, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [SAFE]: The skill provides a structured framework for generating static educational content. It follows best practices for rendering Markdown and code without relying on risky external runtimes or CDNs.\n- [COMMAND_EXECUTION]: Local build and setup scripts (build-hub.mjs, setup-deploy.sh) execute system commands to manage the file system and run the Node.js package manager. These operations are essential to the skill's purpose and are limited to the target project directory. Additional execution of the Mermaid CLI (mmdc) occurs locally for diagram rendering.\n- [EXTERNAL_DOWNLOADS]: The project uses standard Node.js dependency management. Building the learning hub triggers downloads of common development tools and libraries from the npm registry, such as Vite, React, and various Markdown processing utilities.\n- [PROMPT_INJECTION]: As the skill processes user-supplied and external repository data, there is a surface for indirect prompt injection (Category 8). The skill proactively mitigates this via a mandatory evidence chain:\n
- Ingestion points: User-provided Markdown DSL files and external repository/PR data ingested via recipes (references/source-recipes.md).\n
- Boundary markers: Use of specific container delimiters (:::quiz, :::reveal) and a hard "Audience principle" policy to separate meta-commentary from content.\n
- Capability inventory: Local file writing in scripts/build-walkthrough.mjs and automated scaffolding in scripts/setup-deploy.sh.\n
- Sanitization: A specialized "audience linter" (AUDIENCE_RULES) in the build engine identifies and flags author-facing instructions or AI-generated meta-commentary for removal, ensuring the final artifact remains focused on the learner.
Audit Metadata