workflow-creator

Workflow Creator

Guide users to create standardized, executable Talos workflows starting from intent.

Produce a set of files:workflows/<name>/ in the directory workflow.md and manifest.json, installable into the target project via talos install <name> .

Architecture

Layer	File	Responsibility
Orchestration	workflow.md	Define stages and coordination rules
Configuration	manifest.json	Declare dependencies (agents, skills, mcp, plugins)
Agent	agents/*.md（root）	Reusable builtin agent library, referenced by workflows via paths
Memory	memorizer skill	Guided loading by the coordinator when delegating to subagents

Decoupling Principles：

1. workflow.md Only references agent names (委托 **X** agent), does not embed agent logic
2. Agents do not hardcode skill content, only reference skill names (加载 **X** skill）
3. manifest.json is the single source of configuration—dependencies and configuration are all declared here
4. Memory writing is guided by the workflow coordinator via prompt when delegating to subagents, no need to define a stage in workflow.md
5. Each dependency is declared in only one place

Interaction Flow

Phase 1: Capture Intent

Extract from conversation context, ask for what's missing. Clarify four things:

1. Goal: What to accomplish?
2. Starting Point: What is the typical input? (issue, design mockup, prompt?)
3. End Point: What is the expected final deliverable? (PR, deployment, documentation?)
4. Reference: Are there existing workflows to draw from?

Skip redundant questions when intent is already clear.

Phase 2: Design Stages

Work backward from the end point; each precondition is a stage.No need to manually add memory stages (reading and writing memory is handled automatically by the workflow skill).

Division of Labor Decision：

Characteristic	Handling Method	workflow.md Writing
Frequent user interaction (clarification, confirmation, selection)	Main agent handles it itself	No subagent/skill keywords
Well-scoped, self-contained execution (writing code, review, testing)	Delegate to subagent	委托 **X** agent
Reusable procedural patterns (TDD, debugging, PRD)	Load skill	加载 **X** skill

Agent Reuse: When designing stages, first check agents/ for existing builtin agents in the directory. Only suggest creating a new one when no suitable builtin agent exists (place it in the agents/ directory, referenced via manifest.json ).

Output the stage list to the user for confirmation, annotating the execution method in parentheses:

Stage 1: 理解缺陷 [main]
Stage 2: 调试验证 [agent: debugger]
Stage 3: 代码审查 [agent: reviewer]

After user confirmation, proceed to Phase 3.

Phase 3: Skill & MCP Discovery

For each stage that requires external capabilities:

1. User already knows → record directly into manifest.json
2. User describes the requirement but does not know its name → Use WebSearch to search for site:skills.sh <关键词> or guide browsing https://www.skills.sh
3. If not found → Inform the user, and later use /skill-creator to create

Phase 4: Generate files

Generate in order:

1. workflows/<name>/workflow.md
2. workflows/<name>/manifest.json
3. If a new agent is needed:agents/<name>.md

After generation, execute the verification checklist.

Output format

workflow.md

Compatible with workflow runner parsing rules:

• Stage title:## stage N — name
• Subagent delegation:委托 **X** agent
• Skill loading:加载 **X** skill
• Skip condition (optional):XX时跳过

Not needed Include memory-related stages (reading memory and writing memory are handled automatically by the workflow skill).

# <Workflow 名称>

一句话描述。

## stage 1 — <stage 名称>

<stage 描述>

委托 **<agent-name>** agent，传入 <参数>。

<可选：XX时跳过>

manifest.json

{
  "agents": [
    "agents/<builtin-agent-name>",
    "./agents/<local-agent-name>"
  ],
  "skills": [
    { "name": "<name>", "source": "<source>" }
  ],
  "mcp": [
    {
      "name": "<name>",
      "transport": "stdio",
      "command": "<command>",
      "args": ["<arg>"]
    }
  ],
  "plugins": [
    "<plugin-ref>"
  ]
}

Field description:

• agents: Path reference.agents/xx references builtin (root agents/),./agents/xx references workflow-local
• skills: Download from the skills.sh registry
• mcp: Inline configuration object, or path reference (./mcp/config.json）
• plugins: plugin ref string

agents/*.md

Only create a new agent when there is no suitable builtin agent, place it in the root agents/ directory:

---
name: "<kebab-case>"
description: "<一句话：做什么 + 何时用>"
tools: Bash, Read, Write, Edit, Skill
model: sonnet
---

<2-3 句职责说明>

## 输入

<接收什么>

## 输出

<产出什么>

## 流程

1. <步骤>
2. ...

## 依赖

- **Skills**: `<name>` — <使用条件>
- **MCP**: <server> — <用途>

## 规则

- <约束>

Verification checklist

1. For each X of 委托 **X** agent in workflow.md, in manifest.json there is a corresponding path in agents
2. For each skill in agent ## 依赖 , there is an entry in manifest.json 's skills
3. For each MCP in agent ## 依赖 , it is declared in manifest.json 's mcp
4. manifest.json has no orphan entries (all declared dependencies are referenced by workflow.md or an agent)
5. Stage numbers increase consecutively starting from 1
6. The completion criteria of each stage are clearly determinable in the description
7. workflow.md does not contain memory-related stages (handled automatically by the workflow skill)

Complete example: Debug Pipeline

workflow.md：

# Debug Pipeline

Web 缺陷诊断与修复流程。

## stage 1 — 理解缺陷

与用户对话，理解缺陷现象、复现步骤、期望行为。

在完整理解缺陷之前可以多次询问用户，不要急于进入调试阶段。确认理解后输出结构化的缺陷描述。

## stage 2 — 调试验证

委托 **debugger** agent，传入缺陷描述和相关代码位置。

## stage 3 — 代码审查

委托 **reviewer** agent，传入修复代码。

缺陷为单行配置变更或样式微调时跳过。

manifest.json：

{
  "agents": [
    "agents/debugger",
    "agents/reviewer"
  ],
  "skills": [
    { "name": "diagnose", "source": "mattpocock/skills" },
    { "name": "code-review", "source": "obra/superpowers", "installName": "requesting-code-review" },
    { "name": "knowledge-synthesis", "source": "anthropics/knowledge-work-plugins" },
    { "name": "obsidian-markdown", "source": "kepano/obsidian-skills" }
  ],
  "mcp": [],
  "plugins": []
}

agents/debugger.md：

---
name: "debugger"
description: "诊断并修复缺陷。当需要分析 bug、定位根因、修复问题并验证时使用。"
tools: Bash, Read, Write, Edit, Skill
model: sonnet
---

循环诊断缺陷：用 diagnose skill 深度分析，根据环境选择合适的工具验证并采集 debug 数据，直到问题确认解决。

## 输入

- 缺陷描述和复现步骤
- 相关代码位置（如有）

## 输出

- 根因分析
- 修复后的代码
- 验证结果

## 流程

1. 加载 `diagnose` skill 进行系统化分析，形成初始假设
2. 根据缺陷环境选择工具验证假设（浏览器用 Chrome DevTools，后端用日志/断点等）
3. 根据收集的数据修正假设，定位根因
4. 实现修复
5. 再次验证修复效果
6. 验证不通过则回到步骤 1，带入新数据继续分析
7. 验证通过后输出结论

## 依赖

- **Skills**: `diagnose` — 系统化 bug 诊断

## 规则

- 每轮循环必须收集新的证据，不允许空转
- 优先复现再修复，不在没有证据的情况下改代码
- 修复后必须验证原场景通过