The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The "agents/launch-agent.md" file instructs the agent to access and read ".env" files to retrieve Personal Access Tokens ("USER_PAT" and "LINK_PLATFORM_PAT"). Accessing sensitive file paths containing credentials is a high-risk activity.
[DYNAMIC_EXECUTION]: The script "scripts/generate-project.js" utilizes "execSync" to dynamically assemble and execute command strings based on the "techStack" variable. This pattern is often used to execute code that is not statically verifiable.
[COMMAND_EXECUTION]: The skill relies on the "bash" tool to execute a suite of local Node.js scripts ("init-project.js", "generate-html.js", etc.) for its core functionality.
[DATA_EXFILTRATION]: The "launch-agent.md" contains logic to send generated site data and collected credentials to external domains ("10x.in") via the "site-deployments" API.
[PROMPT_INJECTION]: The "SKILL.md" contains strict behavioral constraints and overrides (e.g., "NEVER mention any external courses") designed to steer the agent's output regardless of context.
[EXTERNAL_DOWNLOADS]: The skill automatically injects external script references into generated pages, specifically the WebMCP library from "https://cdn.jsdelivr.net/npm/webmcp@latest/webmcp.js". While this is a well-known service, it introduces a third-party dependency into the final output.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates a high vulnerability to indirect prompt injection. Ingestion points: User requirements collected via "ask-user" and stored in "user-preferences/{project}.json". Boundary markers: None identified in the prompt templates for the specialist agents. Capability inventory: File system writes ("write" tool), command execution ("bash" tool), and network exfiltration via API ("launch-agent.md"). Sanitization: There is no evidence of sanitization or escaping of user-provided content before it is interpolated into agent prompts or generated HTML/JS code.

landing-page