The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and integrate data from untrusted external sources (URLs and pasted text) into a local file-based knowledge base, creating a surface for indirect prompt injection.
Ingestion points: The skill accepts external content in Step 2 ("Ingest") of the workflow defined in SKILL.md, supporting URLs, paths, and pasted text.
Boundary markers: The agent is instructed to follow strict contracts defined in reference/distillation.md and reference/citation.md. These files provide structured rules for how insights should be integrated and cited, which acts as a guide for the agent when processing potentially malicious external instructions.
Capability inventory: The orchestrator utilizes the Agent tool to dispatch subagents and the Skill tool to perform repository-wide operations including file writes and updates in the knowledge/ and sources/ directories.
Sanitization: The workflow includes specific verification (kb-verify) and repair (kb-repair) phases designed to audit the integrity of the knowledge base and fix defects or drifted content, though it lacks explicit input sanitization for the ingested raw text.

kb-builder