AgentSkillsCN

architecture-review

依据明确的RAG架构文档,对现有的Python RAG代码库进行审查,并制定一份包含优先级、设计依据与执行指引的生产就绪待办列表。

SKILL.md
--- frontmatter
name: architecture-review
description: Review an existing Python RAG codebase against an explicit RAG architecture document and produce a production-readiness backlog with priorities, rationale, and execution guidance.

Architecture Review Skill

Goal

Produce a production-grade architecture review document for a non-trivial Python RAG system.

The output must be a prioritised, implementation-ready task backlog that a senior engineering team can directly execute to harden the system for production.

The review must explicitly evaluate:

  1. The current implementation
  2. The stated architecture document (RAG_ARCHITECTURE.md)
  3. Known, explicitly deferred architectural areas

When to Use This Skill

Use this skill when the user asks to:

  • Review a RAG or LLM system for production readiness
  • Audit architecture compliance
  • Identify gaps, risks, or missing safeguards
  • Produce a structured backlog for engineering execution

Instructions

  1. Assume the system is non-trivial and already deployed in at least one environment.
  2. Assume the audience is senior engineers.
  3. Evaluate the codebase against:
    • Production engineering standards
    • The architecture document
    • Explicit architectural constraints (including deferred areas)
  4. Produce one table per category, in the exact order specified below.
  5. Each task must be concrete, opinionated, and actionable.
  6. Do not explain code line-by-line.
  7. Focus on what must be done, not how the system generally works.

Categories

(Create EXACTLY one table per category, in this order)

  1. Architecture & System Design
  2. RAG Pipeline (Ingestion, Retrieval, Reranking, Generation)
  3. Prompting & Grounding Controls
  4. Code Quality & Python Practices
  5. Observability & Operations
  6. Evaluation & Quality Measurement
  7. Security, Privacy & Data Risks
  8. Performance, Latency & Cost
  9. Testing Strategy

Table Format (MANDATORY)

Each table must contain EXACTLY the following columns, in this order: