The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, creating a potential surface for indirect prompt injection attacks.
Ingestion points: The skill ingests data from the company_description field and performs external fetches of website content via a rendering proxy defined in clay-table.md (Column 4) and clay-workflow.md (Node 4).
Boundary markers: The LLM prompt in SKILL.md (Section 6) includes explicit constraints (Rules 5, 9, and 10) that instruct the model not to invent facts, to avoid generic phrases, and to return empty values if the input data is insufficient.
Capability inventory: The skill utilizes text generation capabilities via gpt-4o-mini. It does not have access to sensitive file writing or arbitrary system command execution.
Sanitization: A robust deterministic guard is implemented in both JavaScript (clay-table.md, Column 7) and Python (clay-workflow.md, Node 6) to verify that generated content is grounded in the provided evidence and adheres to a strict whitelist defined in the client_offer_block.
[EXTERNAL_DOWNLOADS]: The skill performs automated network requests to fetch external website content for enrichment purposes.
Evidence: The workflow includes a rendering proxy step (clay-table.md Column 4 and clay-workflow.md Node 4) designed to retrieve text from prospect homepages. These requests are targeted at the domains provided in the input and are a standard part of the skill's documented functionality.

playbook-ai-specificity