The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches text from external, untrusted websites (company homepages and team pages) to perform lead qualification using an LLM. This represents an indirect prompt injection attack surface.
Ingestion points: Website text is scraped in scripts/classify-batch.ts, scripts/multi-lane-classify.ts, and scripts/solo-teampage.ts.
Boundary markers: The prompts use --- and EVIDENCE: delimiters to separate external content from instructions.
Capability inventory: The LLM's classification determines whether domains are qualified, rescued from rejection, or enriched with further data.
Sanitization: Content is truncated to 3,200 or 12,000 characters but is not explicitly sanitized for embedded instructions.
[COMMAND_EXECUTION]: The orchestrator scripts (run-lane.ts, snowball.ts, contacts.ts) use spawn and execSync to manage a multi-stage pipeline by executing subordinate scripts within the same package.
Scripts executed: Includes lookalikes.ts, pull.ts, score-batch.ts, verify-website.ts, enrich-domains.ts, contact-count.ts, and push-sheet.py.
Safety check: These are local scripts identified by path and are necessary for the skill's complex operational model.
[DATA_EXPOSURE]: The skill retrieves API keys from the local environment and specifically reads the platform configuration file at ~/.claude.json to find credentials for the GetLeads service.
File access: scripts/getleads-client.ts reads ~/.claude.json to extract Authorization headers.
[REMOTE_SERVICE_INTERACTION]: The skill interacts with several external APIs for its core functionality.
Services: api.openai.com (qualification), api.prospeo.io (discovery), app.getleads.io (contact counts), and Google Sheets API (delivery). All are well-known services relevant to the skill's purpose.

list-builder