The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting content from live websites and using it to influence AI-driven company qualification decisions.
Ingestion points: Untrusted data is retrieved from external homepages using the fetchHomepageText function in scripts/lib.ts, which is utilized by scripts/score-batch.ts and scripts/verify-website.ts to qualify companies.
Boundary markers: The prompts in the scoring scripts use basic delimiters like '---' and labels such as 'EVIDENCE:'. These markers provide basic separation but are not robust against adversarial content embedded in the web data.
Capability inventory: The skill has the capability to write data to the local filesystem (reports and CSVs) and communicate with external APIs including OpenAI, Prospeo, Exa, and Parallel.ai.
Sanitization: External content is truncated to 3,500 characters, but no further sanitization, filtering, or escaping is performed before interpolation into the AI prompts.
[COMMAND_EXECUTION]: The skill pipeline relies on executing multiple local TypeScript scripts using the tsx runner. While these scripts are part of the skill's own package, the workflow involves the agent performing file system operations and network requests via shell commands.

list-expander