fredrikc3-sketch/vibeconfigstarterkit

Dynamics 365 财务与运营

实施加速器 — 技能版

一个由人工智能驱动的工作空间，用于配置、部署和文档化Dynamics 365财务与运营环境——基于Anthropic Skills架构和官方Dynamics 365 ERP MCP服务器。

---

这是什么

一个结构化、AI可读的知识库，将D365 F&O实现从顾问记忆+零散的电子表格转变为可重复、自我改进的流程。每项能力都存在于技能中——一个带有YAML前言的小型SKILL.md，由AI按需加载。

为什么是技能？

传统的副驾驶指令文件将所有内容导入始终加载的基础上下文中。技能使用渐进披露：只有每个技能的description是永久可见的。只有当任务匹配时，整个身体才加载。结果：

• 更小且始终加载的上下文 →更快、更便宜的转弯。
• 每个能力独立演进——没有单一指令文件会无限增长。
• 行业标准、模型无关格式（人类技能）。

架构：四层编排

v2的重大变化。

Dynamics 365Finance & Operations

Implementation Accelerator — Skills Edition

An AI-driven workspace for configuring, deploying, and documenting Dynamics 365 Finance & Operations environments — built on the Anthropic Skills architecture and the official Dynamics 365 ERP MCP server.

---

What this is

A structured, AI-readable knowledge base that turns D365 F&O implementations from consultant memory + scattered spreadsheets into a repeatable, self-improving process. Every capability lives in a Skill — a small SKILL.md with YAML frontmatter that the AI loads on demand.

Why Skills?

Traditional copilot instruction files dump everything into the always-loaded base context. Skills use progressive disclosure: only each skill’s description is permanently visible. The full body loads only when the task matches. The result:

• Smaller always-loaded context → faster, cheaper turns.
• Each capability evolves independently — no monolithic instruction file to grow forever.
• Industry-standard, model-agnostic format (Anthropic Skills).

Architecture: 4-Layer Orchestration

The big change in v2.1 is multi-layer orchestration with sub-agent delegation. Per-module work — config building, deployment, validation — is dispatched to sub-agents that load only their own module’s knowledge. The main agent’s context stays small even on a project with 25+ modules.

v2.2 council-driven hardening layers on top:

• Canonical JSON Schemas in schemas/ for journal entries, DMF templates, run-state, dependency graph and the fan-out worker contract — enforced by tools/validate_repo.py and a CI quality gate.
• Single-source-of-truth dependency graph at Modules/dependency-graph.json. Skills derive waves from this file, no more hand-rolled wave lists.
• Concurrent-write safety: workers write per-module artefacts under Documentation/_status//.json; the orchestrator merges into shared MD files after each wave.