The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages the vendor-specific 'postplus' CLI to perform research tasks. These commands are executed locally to fetch TikTok metadata, comments, and profile information. The skill includes a robust cost-management workflow where the agent must obtain explicit user approval for 'quote-confirmation challenges' before proceeding with potentially paid operations.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze third-party content from TikTok, such as video comments and captions, which represents a surface for indirect prompt injection.
Ingestion points: External data enters the agent's context through CLI routes like 'tiktok-comments' and 'tiktok-videos' as described in references/audience-voice.md and references/organic-benchmark.md.
Boundary markers: The instructions provide guidelines to identify 'low-signal' or 'spam-heavy' comments, and they mandate that the agent report evidence quality and limits, which serves as a logical boundary between data and analysis.
Capability inventory: The agent uses CLI tools to read from the network and write results to local JSON files (result.json).
Sanitization: No explicit string sanitization is mentioned, but the workflow emphasizes high-level categorization (objections, FAQ, praise) which mitigates the risk of the agent executing embedded instructions.
[SAFE]: The skill follows security best practices by explicitly instructing the agent not to ask for credentials, private exports, retry strategies, or cookies. It also clearly defines the boundary between public research and private/backend data (Shop, LIVE, analytics), instructing the agent to stop if the user requests access to private surfaces.

tiktok-research