Title: Agent Skill Catalog Version: 1.1 Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] Status: Draft (target-state catalog; most rows are unimplemented — see Status column) Last reviewed: 2026-09-08 Next review: [TENANT_CONFIGURATION_REQUIRED] Reviewers: AI Governance, Architecture
Catalogs every specialist skill planned for the platform (agent-architecture.md), its inputs/outputs, tool access, and confidence threshold. This is a target catalog, not a list of what’s live — see the Status column below; only 3 of 7 rows have real backing logic today. Each skill has a corresponding system prompt in prompts/system/, but a prompt file existing does not imply the skill is implemented (prompts were authored independently of the backend build-out).
| Skill | Status | Purpose | Input | Output (JSON-Schema validated) | Tools (least privilege) | Confidence threshold | Human-review trigger |
|---|---|---|---|---|---|---|---|
| CV Extraction | Implemented (parse_cv_skill, HrAutomation.Agents/CvIngestion/) |
Parse CV into structured fields | Raw CV (untrusted) | Structured candidate fields + confidence per field | None (pure NLP task) | 0.75 | Below threshold, or dedup flag |
| Candidate Matching | Planned (stub only — always returns blocked) |
Score candidates against approved JD | Approved JD criteria + CV Bank subset | Ranked list with per-candidate rationale | Read-only CV Bank query tool | 0.60 (score inclusion) | Any candidate below threshold excluded, not force-included |
| Interview Coordination Drafting | Planned (stub only) | Suggest interview slots, draft invite text | Panelist calendars (via MCP, read-only) | Proposed slot list + draft invite text | mcp-calendar (read-only availability) |
N/A (proposal only, no threshold) | Any conflict/no-availability case routed to recruiter |
| Offer Drafting | Planned (stub only) | Draft offer letter from template + compensation reference | Approved template + compensation reference (external) | Draft offer document (no compensation figure authored by AI — reference only) | None (template rendering only) | N/A | Always human-approved before send (mandatory gate) |
| Document Verification Assist | Planned (stub only) | Cross-check submitted documents against Green Form data | Document OCR/extraction + self-reported data | Match/mismatch flags with confidence | Document OCR tool (internal) | 0.80 | Below threshold → needs_review, never auto-pass |
| Employee Conversion Assist | Planned (stub only) | Summarize gate-checklist status for approver | Gate-checklist data (read-only) | Summary of pass/fail per check | Read-only gate-checklist query | N/A | N/A — decision always human |
| Policy Q&A (RAG) | Planned (no code at all — HrAutomation.Rag is an empty scaffold) |
Answer HR-policy questions with citations | User question + retrieved chunks | Answer + citations, or no-answer fallback | RAG retrieval tool (read-only) | 0.65 (retrieval relevance) | Below threshold → no-answer fallback |
Also implemented but not in the original catalog above: TAN creation (create_tan_skill) and TAN approval/rejection (tan_approval_skill), both in HrAutomation.Agents/Tan/ — added once the real database-first TAN workflow was built. Every “stub only” skill above is registered in HrAutomation.Agents/Stubs/StubSkillCatalog.cs and always returns ActionStatus.blocked with a “not yet implemented” risk note — it never proposes a state transition and never writes to the database.
Every tool grant above is scoped to read-only or propose-only per ADR-004; no skill is granted a write-capable tool to a sensitive external system (e.g., no skill can directly send a calendar invite without the Interview Coordination workflow step confirming with the recruiter first, per mcp-tool-governance.md).
All skills return JSON-Schema-validated output; free-text fields are limited to designated “draft”/”rationale” fields and never interpreted as executable instructions downstream (see prompt-injection-defense.md).
| Version | Date | Author | Change |
|---|---|---|---|
| 1.0 | 2026-09-07 | Documentation package generation | Initial creation |