The Agent Skills Directory

[SAFE]: The skill follows established security best practices for AI agents. It explicitly instructs the agent to establish context without requesting sensitive personal information like names or birthdays. All external resources and taxonomy data originate from the author's own infrastructure (withmarbleapp) or well-known, trusted organizations.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-defined learning goals and external taxonomy data from the 'os-taxonomy' repository.
Ingestion points: User goals provided in the conversation and taxonomy files (topics.json, teaching-research.json) as described in references/data-and-research.md.
Boundary markers: The skill instructions emphasize the use of explicit confirmation checkpoints and 'provisional teaching briefs' when data is missing, reducing the risk of accidental obedience to instructions embedded in data.
Capability inventory: The skill uses esbuild for local bundling and generates standalone HTML. No runtime network operations are used for child-facing experiences by default.
Sanitization: The instructions include guidelines for sanitizing model-generated and child-authored text, particularly in the optional connected LLM mode (references/llm-connected-experiences.md).
[DYNAMIC_EXECUTION]: The skill includes a build script (examples/build.mjs) that uses the standard esbuild library to bundle JavaScript source files into standalone HTML files. This is a primary intended function of the builder skill and operates on local source code within the skill's directory.

marble-builder