搜索

PythonReactSEOPDF
GitHub Skill

agent-workflow

Multi-agent development workflow system. Load when orchestrating development tasks, spawning subagents, or managing workflow phases.

aiskillstore☆ 123 仓库 Stars6 仓库 Forks

安装与使用

请根据来源仓库和你的 AI 工具说明安装此技能。先阅读技能文件和依赖要求,再添加到工作流。

查看来源 ↗使用指南

技能说明

Agent Workflow System

This skill provides the multi-agent workflow orchestration system.

When to Use

Load when:

  • Starting a new development task
  • Spawning subagents for parallel work
  • Managing workflow phase transitions
  • Coordinating multi-agent collaboration

Core Concepts

Workflow Phases
  1. REQUIREMENTS - PM gathers and clarifies user needs
  2. ARCHITECTURE - Architects design solutions in parallel
  3. IMPLEMENTATION - Executor manages parallel implementers
  4. VERIFICATION - QA validates the implementation
  5. REFLECTION - Capture learnings for evolution
Agent Hierarchy

Tier-1 (can spawn subagents):

  • PM - Requirements gathering
  • Architect - Technical design
  • Roundtable - Design synthesis
  • Executor - Implementation orchestration
  • QA - System verification

Tier-2 (focused workers):

  • Explorer - Fast codebase scouting
  • Implementer - Focused code writing
  • Verifier - Independent verification
  • Tester - Test execution
  • Contract Resolver - Handle blocked tasks
  • Reflector - Extract learnings
  • Evolver - System improvement

Agent Definitions

Tier-1 Agents:

  • @.claude/agents/pm.md
  • @.claude/agents/architect.md
  • @.claude/agents/roundtable.md
  • @.claude/agents/executor.md
  • @.claude/agents/qa.md

Tier-2 Agents:

  • @.claude/agents/explorer.md
  • @.claude/agents/implementer.md
  • @.claude/agents/verifier.md
  • @.claude/agents/tester.md
  • @.claude/agents/contract-resolver.md
  • @.claude/agents/reflector.md
  • @.claude/agents/evolver.md

Spawning Pattern

Custom Agents (Dotagent System)

Use lowercase names matching agent definitions:

Task(
  subagent_type: "explorer",    # Fast codebase scouting
  model: "haiku",
  prompt: "Query: ... Output: ..."
)

Task(
  subagent_type: "implementer", # Focused code writing
  model: "sonnet",
  prompt: "Task: ... Boundaries: ... Output: ..."
)

Task(
  subagent_type: "pm",          # Requirements gathering
  model: "sonnet",
  prompt: "Request: ... Output: memory/reports/demand.json"
)
Built-in Agents (Claude Code)

These are different from custom agents above:

Task(
  subagent_type: "Explore",     # Built-in: quick codebase exploration
  model: "haiku",
  prompt: "Find all API endpoints..."
)

Task(
  subagent_type: "Plan",        # Built-in: planning agent
  model: "sonnet",
  prompt: "Plan the implementation..."
)
Subagent Type Reference
Custom (Dotagent) Model Purpose
pm sonnet Requirements gathering
architect opus Technical design
roundtable opus Design synthesis
executor sonnet Implementation orchestration
qa sonnet System verification
explorer haiku Fast codebase scouting
implementer sonnet Focused code writing
verifier haiku Independent verification
tester haiku Test execution
contract-resolver sonnet Handle blocked tasks
reflector sonnet Extract learnings
evolver opus System improvement
Built-in (Claude Code) Purpose
Explore Quick codebase exploration
Plan Planning and analysis
general-purpose Flexible multi-step tasks

Workflow State

@workflows/main.json @memory/state/phase.json

使用示例

查看来源仓库中的使用示例与配置说明。

查看来源 ↗

来源历史

收录时间
2026-01-20
索引更新时间
2026-01-28
Branch
main

收录与同步日期属于本站记录。源码修改历史请查看来源仓库。

GitHub ↗

chat Comments (0)

chat_bubble_outline

No comments yet. Be the first to share your thoughts!