The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from untrusted external URLs using mcp__Claude_in_Chrome__read_page and WebFetch. If a target website contains malicious text or hidden instructions designed to manipulate the agent, it could influence the code generation process.
Ingestion points: Target page accessibility tree (read_page) and raw HTML (WebFetch).
Boundary markers: Absent; there are no specific instructions to ignore malicious directives found within the scraped content.
Capability inventory: File system write (project creation), shell command execution (npm install), and network operations (Chrome automation).
Sanitization: Not explicitly performed on the ingested content before it is used for code generation.
[DYNAMIC_EXECUTION]: The skill dynamically assembles a Next.js project based on design tokens and layout structures extracted from external websites. The generated code is intended to be executed locally by the user.
[COMMAND_EXECUTION]: The instructions direct the user to execute shell commands including npm install, npm run dev, and npx tsc --noEmit on the project folder created by the agent.
[REMOTE_CODE_EXECUTION]: The skill uses mcp__Claude_in_Chrome__javascript_tool to execute JavaScript snippets (documented in references/extraction-scripts.md) on remote web pages provided by the user to facilitate data extraction.
[UNVERIFIABLE_DEPENDENCIES]: The project scaffold in references/project-scaffold.md relies on several external packages (next, react, tailwindcss, etc.) which are downloaded and installed at runtime when the user follows the execution instructions.

pm-url2proto