| Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] |
Status: Draft |
Purpose
Reserved for tool-use instruction fragments specific to invoking MCP tools correctly and safely (e.g., how to format a get_calendar_availability call, how to handle a tool error response). These fragments compose into the relevant skill’s system prompt in prompts/system/ — see interview-coordination-system-prompt.md for the primary consumer today.
Convention
- One file per tool or tool group, named after the MCP server (e.g.,
mcp-calendar-tool-use.md).
- Every fragment must reiterate the tool’s tier (read-only / propose-only / approval-required write per hr-tools.schema.json) so the model never assumes broader capability than granted.
- Tool response content is always treated as untrusted (see prompt-injection-defense.md) — fragments must instruct the model accordingly.
Change control
| Version |
Date |
Author |
Change |
| 1.0 |
2026-09-07 |
Documentation package generation |
Initial creation |