The Agent Skills Directory

[PROMPT_INJECTION]: The 'Reporting' section in SKILL.md (lines 83-93) instructs the agent to perform a self-audit using specific severity levels and a 'Block' or 'Approve' verdict. This mimics security scanning protocols and could be used to mislead reviewers by providing a false sense of security or overriding external evaluation logic.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data to generate and execute code, creating a surface for injection attacks.
Ingestion points: sketchbook/taste.md and project source files (framework, styling, tokens).
Boundary markers: None; the skill does not use delimiters or instructions to ignore embedded commands in the source files.
Capability inventory: File-system write (/sketchbook, sketchbook/kept/), file-system delete (sketchbook/delete-the-killed), and execution of a development server (sketchbook-open-it).
Sanitization: None; project code and preferences are read and used directly to generate new interactive components.
[DYNAMIC_EXECUTION]: The skill is designed to generate interactive code (hover, toggle, drag, etc.) at runtime based on project context and execute it as 'live playable tiles' within a development server environment.

sketchbook