The Agent Skills Directory

Handling of External Evidence: The skill is designed to process external context from sources like meeting transcripts, public filings, and document stores. While processing external data is a potential surface for indirect prompt injection, the skill mitigates this through strict evidence hierarchy rules that prioritize customer-provided facts over external research.
Ingestion points: Data enters the agent context through the crm, meeting_notes, document_store, and public research categories as defined in SKILL.md.
Boundary markers: The skill relies on narrative labeling (e.g., Known, Inferred, Assumed) to distinguish between different types of evidence rather than explicit technical delimiters.
Capability inventory: The skill focus is on data retrieval and analysis; no evidence of subprocess execution, dynamic code evaluation, or unauthorized network operations was found.
Sanitization: The instructions focus on source authority and validation logic rather than technical input sanitization.
Centralized Context Management: The skill implements a mandatory preflight gate that invokes a specific helper skill (sales:user-context) to resolve source categories and context obligations. This design pattern centralizes security-sensitive tasks like source resolution and permissions.
Source Discipline and Provenance: The instructions in SKILL.md and references/value-model-and-evidence.md mandate clear labeling of source provenance and explicitly forbid the use of public strategic language to invent company-specific impact, reducing the risk of misinformation.
No Executable Code: The skill consists entirely of configuration and instructional markdown files. No scripts (Python, JavaScript, shell) or compiled binaries are included, which significantly limits the potential for traditional code-based exploits.

build-business-case