The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze data from several external and untrusted sources, including documentation pages from docs.getdbt.com, dbt Cloud job logs, and git repository contents. This constitutes a potential surface for indirect prompt injection attacks.
Ingestion points: Documentation content (fetched via WebFetch in fetching-dbt-docs), dbt Cloud API responses and run artifacts (in troubleshooting-dbt-job-errors), and project metadata like manifest.json and catalog.json.
Boundary markers: The skill contains robust, explicit instructions across its components (notably in fetching-dbt-docs, using-dbt-for-analytics-engineering, and troubleshooting-dbt-job-errors) commanding the agent to treat all external content as untrusted and to never execute instructions found embedded within such data.
Capability inventory: The agent has access to the dbt CLI, bash shell commands, and file system operations (Read/Write/Edit), which increases the potential impact of an injection if the guardrails were bypassed.
Sanitization: The instructions emphasize the extraction of structured data fields while ignoring any instruction-like text that might attempt to override the agent's behavior.

dbt