メニュー
Use when the user asks for analysis that requires decomposing a complex question into multiple queries (trend analysis, comparisons, aggregation across entities)
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Dispatch ERP queries to Workers. Handles task creation, MinIO sync, state registration, and Matrix room dispatch.
Manager direct-to-MCP fast path for simple ERP data queries. Bypasses Workers entirely — calls honeybadge-nebula-mcp generate_query + validate_and_execute, then forwards results to user via contract 002.
Use when handling any natural language question about ERP data (suppliers, POs, invoices, payments, items, receipts, etc.)
Use when the user asks about fraud detection, three-way matching anomalies, duplicate invoices, unusual payment patterns, or supplier concentration risk
| name | multi-step-analysis |
| description | Use when the user asks for analysis that requires decomposing a complex question into multiple queries (trend analysis, comparisons, aggregation across entities) |
Call the Python analysis modules:
# Decompose a complex question into sub-queries
python3 -m multi_step_analysis.lib.decompose --question "对比2025年和2026年Q1的采购金额变化"
# Cross-reference results from multiple rounds
python3 -m multi_step_analysis.lib.decompose cross-reference \
--results-dir /tmp/mcp_results/
python3 -m multi_step_analysis.lib.decompose --question "<QUESTION>"
Breaks the complex question into 2-5 sub-queries with round numbers.
For each sub-query: use common.mcp_client to generate_query -> validate_and_execute.
python3 -m multi_step_analysis.lib.decompose cross-reference --results-dir /tmp/
Finds patterns, trends, or anomalies across sub-query results.
Present findings with severity levels:
CRITICAL: All numbers must come directly from query results. Do NOT calculate values not in the database.
Write one audit log entry capturing all sub-queries and the final analysis.