The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted lead data from external sources and interpolates it into prompts.
Ingestion points: scripts/classify_leads_llm.py reads company information (name, description, industry, keywords) from input JSON files, while scripts/read_sheet.py ingests data directly from Google Sheets.
Boundary markers: The prompt templates in scripts/classify_leads_llm.py do not use delimiters (like XML tags, triple backticks, or specific markers) to wrap external data, making it easier for malicious content in a lead's description to influence the LLM's output.
Capability inventory: The skill possesses file system write access (saving classified results) and network access to Anthropic's Batch API and Google's Sheets/Drive APIs.
Sanitization: While the script truncates description and keyword fields, it does not sanitize them for instruction-like text or control characters before prompt construction.
[CREDENTIALS_UNSAFE]: The skill manages sensitive credentials including ANTHROPIC_API_KEY and Google OAuth tokens. scripts/read_sheet.py and scripts/update_sheet.py generate and store token.json in the local directory. While this follows standard OAuth library patterns, users should be aware that these files contain sensitive access tokens.

classify-leads