The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external web sources (search results and page text) to create citations and verify evidence. This ingestion of untrusted content represents an attack surface for indirect prompt injection.
Ingestion points: The scripts/sources.py script ingests URLs and page text via the ingest and quote subcommands.
Boundary markers: The skill uses bracketed numeric identifiers (e.g., [1]) and structured Sources: headers to distinguish cited content from prose.
Capability inventory: The script has capabilities to write to a local JSON ledger and modify draft markdown documents using the --replace-in flag.
Sanitization: The attach_quote function implements a verbatim matching check between proposed quotes and fetched evidence text, normalizing whitespace and markdown characters to ensure claims are grounded in the actual source content.

grounded-citations