The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided data (pasted LinkedIn posts) to generate a voice profile file. This surface is vulnerable to indirect prompt injection where an attacker could include malicious instructions in their posts to influence how the agent behaves or how other skills process the resulting voice file.
Ingestion points: Processes three to five user-pasted posts or messages in the Process section.
Boundary markers: No specific delimiters or "ignore instructions" warnings are defined for the input samples.
Capability inventory: The skill writes to the file system (~/.liftli/voice.md) and invokes tools like strategist_briefing.
Sanitization: No sanitization or filtering of the user-provided samples is performed before they are analyzed and quoted in the output file.
[EXTERNAL_DOWNLOADS]: The skill directs users to fetch configuration and tools from the developer's infrastructure, specifically https://mcp.liftli.ai/mcp and https://liftli.ai/llms.txt.
[COMMAND_EXECUTION]: The skill provides the user with shell commands to install and authenticate with external MCP services (claude mcp add and claude mcp login). While these are intended for platform integration, they involve the execution of setup commands and the addition of remote transport layers to the agent environment.

linkedin-voice