The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from various sources, including CRM exports, uploaded PDFs, meeting transcripts, and user-provided notes as described in the workflow steps and input rules.
Ingestion points: Untrusted data enters the agent context through file uploads (PDFs), CRM data pastes, and external web content fetched via search tools (referenced in SKILL.md Step 1 and references/input-rules.md Rule 5).
Boundary markers: The skill implements strong boundary markers. It contains a mandatory 'Untrusted content' block in SKILL.md (Section 0) that explicitly instructs the agent to treat fetched or uploaded text as data, never as instructions, and requires the agent to report any attempted instruction overrides.
Capability inventory: The skill utilizes web_search, web_fetch, and file generation tools (e.g., creating PPTX files).
Sanitization: The skill relies on clear instructional guardrails and the 'One Dump' principle to normalize diverse inputs, while maintaining a 'Question Budget' to clarify ambiguities.

qbr-prep