The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses LLM agents (Claygent) to process content from various public web sources (SEC filings, news sites, LinkedIn, press releases), which represents a surface for indirect prompt injection attacks where malicious instructions could be embedded in the scraped data.
Ingestion points: Nodes 4 through 7 in the workflow (defined in SKILL.md and references/prompts.md) execute web searches and process external content.
Boundary markers: The prompts include specific instructions for output structure (JSON) and data validation (requiring dates and source URLs), which helps constrain model behavior, but they do not use explicit delimiters for the external content itself.
Capability inventory: The skill has the capability to write research summaries to the user's conversation, CSV files, and Markdown files, and it manages workflows within the Clay platform.
Sanitization: Step 5 of the skill logic includes sanitization rules that drop events with invalid dates or specific characters (e.g., '@' in date fields).
[DYNAMIC_CONTEXT_INJECTION]: The skill uses the ! command syntax in SKILL.md to execute shell commands at load time for environment verification.
Evidence: The skill executes !clay whoami; echo "exit_code=$?" to check CLI status and workspace identity before proceeding. These are benign diagnostic commands using the vendor's official CLI tool.
[COMMAND_EXECUTION]: The skill instructions involve extensive use of the clay CLI to list, create, and manage workflows.
Evidence: Commands such as clay workflows list, clay workflows graph get, and clay workflows runs test are used throughout the setup and execution process. These operations are consistent with the skill's primary purpose and are performed using the vendor's own tooling.
[DATA_EXPOSURE]: The skill includes instructions for managing configuration data that adhere to security best practices.
Evidence: In the 'Declared inputs' section, the skill specifies that when saving answers to a file, it should save "identifiers only — never a token or a password," reducing the risk of accidental credential exposure.

find-recent-company-developments