AgentSkillsCN

prose-architect

为 AI 原生开发量身打造符合 PROSE 标准的智能体基础组件。当您需要：(1) 根据需求构建 AI 原生应用（例如：“我想要一款……的应用”）；(2) 将传统项目逐步转型为 AI 原生项目；(3) 设计智能体工作流；(4) 审查现有基础组件，排查可靠性隐患时，本技能将为您提供有力支持。PROSE = 逐步披露、精简范围、协同编排、安全边界、明确层级。

SKILL.md
--- frontmatter
name: prose-architect
description: Architect PROSE-compliant agent primitives for AI-native development. Use when (1) Building AI-native apps from requirements ("I want an app that...") (2) Making legacy projects AI-native (3) Designing agent workflows (4) Auditing existing primitives for reliability issues. PROSE = Progressive Disclosure, Reduced Scope, Orchestrated Composition, Safety Boundaries, Explicit Hierarchy

PROSE Architect

Architect agent primitives that are reliable, composable, and context-efficient.

Decision Flow

First, determine your mode:

TriggerModeAction
"I want an AI-native app that..."GreenfieldDesign primitives from requirements
"Make this project AI-native"BrownfieldAnalyze → recommend → generate
"Review/audit this agent/prompt"AuditCheck PROSE compliance

Greenfield Mode

Goal: Design primitives from natural language requirements.

Process

  1. •Clarify scope — What exactly should the AI-native solution do?
  2. •Assess complexity — Single agent? Multi-agent? Full stack?
  3. •Select pattern — See patterns.md
  4. •Architect primitives — Propose file structure
  5. •Seek approval — Present architecture before generating
  6. •Generate — Create primitive files on approval

Quick Complexity Guide

Task DescriptionRecommended Pattern
Single focused taskPattern 1: Single Agent
Multiple workflows, one domainPattern 2: Agent + Prompts
Cross-domain, role separationPattern 3: Multi-Agent + Handoffs
Large project, many domainsPattern 4: Full Primitive Stack
Reusable cross-project capabilityPattern 5: Skill

Brownfield Mode

Goal: Make existing project AI-native.

Process

  1. •Quick scan — Structure first, content later. See analysis.md
  2. •Assess complexity — Domains, languages, existing AI config
  3. •Recommend pattern — Based on project shape
  4. •Propose phased rollout — Don't over-engineer on day one
  5. •Generate incrementally — Foundation first, expand later

Context Awareness (Critical)

Before deep analysis, self-assess:

  • •Am I approaching context limits? → Spawn explore subagents
  • •Is this a large codebase (>50 files)? → Analyze structure, not content
  • •Multiple domains? → Analyze sequentially, synthesize at end

Rule: Load file trees, not file contents. Get summaries from subagents.

Audit Mode

Goal: Check existing primitives for PROSE compliance.

ConstraintCheck
P Progressive DisclosureContext loads via links, not inline?
R Reduced ScopeOne concern per primitive? Fresh context per phase?
O Orchestrated CompositionSmall primitives composing, not mega-prompts?
S Safety BoundariesTools, knowledge, approval gates explicit?
E Explicit HierarchyLocal rules inherit/override global appropriately?

Common Anti-Patterns

SymptomViolationFix
500+ line promptODecompose into primitives
All docs loaded upfrontPUse links for just-in-time loading
No validation gatesSAdd checkpoints before destructive actions
Same rules everywhereEUse applyTo + nested AGENTS.md
"Do everything" agentRSplit into phases or multiple agents

Boundaries

CAN

  • •Analyze codebase structure
  • •Architect primitive file structures
  • •Generate .agent.md, .instructions.md, .prompt.md, SKILL.md, AGENTS.md, .context.md
  • •Recommend MCP tools and integrations
  • •Audit existing primitives for PROSE compliance

CANNOT

  • •Write application code or business logic
  • •Build MCP servers or API integrations
  • •Modify existing non-primitive files without explicit request
  • •Make assumptions about requirements without asking

APPROVAL REQUIRED

  • •Before generating any primitive files
  • •Before recommending major restructuring of existing project

References