The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from untrusted external websites to generate code and documentation, creating a surface for indirect prompt injection.
Ingestion points: Target website content including text, DOM structure, and computed styles extracted via Chrome MCP in Phase 1 and Phase 3.
Capability inventory: Extensive file system write access to the project directory, network access for asset downloads, and the ability to dispatch sub-agents (builders) to generate code in separate worktrees.
Boundary markers: The instructions lack specific delimiters or "ignore embedded instructions" warnings for the extracted data, which may allow website content to influence the agent's logic or the generated code.
Sanitization: No sanitization or validation of extracted text, CSS attributes, or SVG data is prescribed before they are incorporated into component specifications and source files.
[COMMAND_EXECUTION]: The skill utilizes local command execution for environment setup and build verification.
Evidence: Instructions require running npm run build and npx tsc --noEmit to verify the project state after merges and builder activity.
[DYNAMIC_EXECUTION]: The skill generates and executes scripts at runtime for data extraction and asset management.
Evidence: The workflow specifies using a "recursive walker script in Chrome MCP" for CSS extraction and a script to "enumerate images, videos... then batch-download" assets.

clone-website