dify-dsl-generator

Dify DSL Workflow Generator

A professional Dify workflow DSL/YML file automatic generation tool, based on deep learning from 86+ real workflow cases, capable of automatically generating complete workflow configuration files that comply with the Dify specification according to user business requirements.

Core Features

✅ Complete DSL Generation: Automatically generate complete YML files containing app, dependencies, workflow
✅ Multi-node Support: Supports all node types such as start, llm, answer, code, http-request, if-else, tool, etc.
✅ Intelligent Connections: Automatically generate edges connection relationships between nodes
✅ Parameter Configuration: Intelligently recommend model parameters and prompt configurations
✅ Plugin Integration: Automatically identify and configure required Dify plugin dependencies
✅ Standardized Format: Strictly follow the Dify 0.3.0 version DSL specification

Usage

Basic Usage

生成一个 Dify 工作流用于 [业务需求描述]

Detailed Usage

帮我生成一个 Dify 工作流 DSL 文件:
- 功能: [工作流要实现的功能]
- 输入: [用户输入的内容]
- 处理步骤: [详细的处理逻辑]
- 输出: [期望的输出结果]
- 使用插件: [需要的插件，可选]

Dify DSL File Structure

Based on learning from 86+ real workflow cases, Dify DSL YML files follow the structure below:

1. App Configuration

app:
  description: '工作流描述'
  icon: 🤖
  icon_background: '#FFEAD5'
  mode: advanced-chat  # 或 workflow, agent-chat
  name: 工作流名称
  use_icon_as_answer_icon: false

Mode Description:

advanced-chat: Advanced chat mode (chatflow)
workflow: Workflow mode
agent-chat: AI Agent mode

2. Dependencies

dependencies:
- current_identifier: null
  type: marketplace
  value:
    marketplace_plugin_unique_identifier: 插件唯一标识符

Common Plugins:

langgenius/openai_api_compatible: OpenAI-compatible interface
bowenliang123/md_exporter: Markdown exporter
Other marketplace plugins are added as needed

3. Workflow

kind: app
version: 0.3.0
workflow:
  conversation_variables: []
  environment_variables: []
  features:
    file_upload:
      enabled: false
    speech_to_text:
      enabled: false
    text_to_speech:
      enabled: false
  graph:
    edges: []
    nodes: []

Node Type Details

Start Node

- data:
    desc: ''
    title: 开始
    type: start
    variables:
    - label: 用户输入
      max_length: 1000
      options: []
      required: true
      type: paragraph  # 或 text-input, select, file
      variable: query
  id: 'start'
  position:
    x: 100
    y: 300
  type: custom
  width: 244
  height: 90

Variable Types:

paragraph: Paragraph text (multi-line)
text-input: Single-line text
select: Dropdown selection
file: File upload
number: Number

LLM Large Language Model Node

- data:
    context:
      enabled: false
      variable_selector: []
    model:
      completion_params:
        temperature: 0.7
        max_tokens: 2000
      mode: chat
      name: gpt-4
      provider: openai
    prompt_template:
    - id: 唯一ID
      role: system
      text: 系统提示词
    - id: 唯一ID
      role: user
      text: 用户提示词 {{#变量引用#}}
    title: LLM节点
    type: llm
    vision:
      enabled: false
  id: '节点ID'
  position:
    x: 400
    y: 300
  type: custom

Common model providers:

openai: OpenAI
langgenius/openai_api_compatible/openai_api_compatible: Compatible interface
anthropic: Claude
alibaba: Tongyi Qianwen

Variable reference format:

{{#节点ID.输出变量#}}: Reference the output of other nodes
{{#sys.query#}}: Reference system variables (user input)
{{#节点ID.text#}}: Reference LLM output text

Code Execution Node

- data:
    code: |
      import json

      def main(arg1: str, arg2: str) -> dict:
          # 处理逻辑
          result = process(arg1, arg2)
          return {
              "result": result,
              "status": "success"
          }
    code_language: python3
    outputs:
      result:
        type: string
      status:
        type: string
    title: 代码执行
    type: code
    variables:
    - value_selector:
      - '前置节点ID'
      - 输出变量
      variable: arg1
  id: '节点ID'
  position:
    x: 700
    y: 300
  type: custom

Code Language:

python3: Python 3
javascript: JavaScript (supported in some versions)

Output Types:

string: String
number: Number
object: Object
array[string]: String array
array[number]: Number array
array[object]: Object array

HTTP Request Node

- data:
    authorization:
      config: null
      type: no-auth
    body:
      data: '{"key": "{{#变量#}}"}'
      type: json
    headers: ''
    method: post
    timeout:
      max_connect_timeout: 0
      max_read_timeout: 0
      max_write_timeout: 0
    title: HTTP请求
    type: http-request
    url: https://api.example.com/endpoint
  id: '节点ID'
  position:
    x: 1000
    y: 300
  type: custom

HTTP Methods:

get: GET request
post: POST request
put: PUT request
patch: PATCH request
delete: DELETE request

Authentication Types:

no-auth: No authentication
api-key: API Key
bearer: Bearer Token

If-Else Conditional Judgement Node

- data:
    cases:
    - case_id: case1
      conditions:
      - comparison_operator: contains
        id: 条件ID
        value: 期望值
        variable_selector:
        - '节点ID'
        - 变量名
      id: case1
      logical_operator: and
    logical_operator: or
    title: 条件判断
    type: if-else
  id: '节点ID'
  position:
    x: 1300
    y: 300
  type: custom

Comparison Operators:

contains: Contains
not contains: Does not contain
is: Equals
is not: Not equals
empty: Is empty
not empty: Is not empty

Logical Operators:

and: And
or: Or

Tool Node

- data:
    provider_id: 工具提供者ID
    provider_name: 工具提供者名称
    provider_type: builtin  # 或 api
    title: 工具调用
    tool_configurations: {}
    tool_label: 工具标签
    tool_name: 工具名称
    tool_parameters:
      参数名:
        type: mixed
        value: '{{#变量#}}'
    type: tool
  id: '节点ID'
  position:
    x: 1600
    y: 300
  type: custom

Tool Types:

builtin: Built-in tools (such as search, weather, etc.)
api: API tools
plugin: Plugin tools

Answer direct reply node

- data:
    answer: |
      {{#LLM节点ID.text#}}

      {{#代码节点ID.result#}}
    title: 直接回复
    type: answer
    variables: []
  id: answer
  position:
    x: 1900
    y: 300
  type: custom

Variable Aggregator node

- data:
    advanced_settings: null
    desc: ''
    groups:
    - group_name: 分组1
      output_type: string
      variables:
      - value_selector:
        - '节点ID'
        - 变量名
        variable: 输出变量名
    title: 变量聚合器
    type: variable-aggregator
  id: '节点ID'
  position:
    x: 2200
    y: 300
  type: custom

Parameter Extractor node

- data:
    instruction: 提取指令说明
    is_array: false
    model:
      completion_params: {}
      mode: chat
      name: gpt-4
      provider: openai
    parameters:
    - description: 参数描述
      name: 参数名
      required: true
      type: string
    query:
    - role: user
      text: '{{#输入变量#}}'
    reasoning_mode: prompt
    title: 参数提取
    type: parameter-extractor
  id: '节点ID'
  position:
    x: 2500
    y: 300
  type: custom

Edges connection relationships

edges:
- data:
    isInIteration: false
    isInLoop: false
    sourceType: start
    targetType: llm
  id: 源节点ID-source-目标节点ID-target
  source: '源节点ID'
  sourceHandle: source
  target: '目标节点ID'
  targetHandle: target
  type: custom
  zIndex: 0

Connection rules:

Each node has at most one incoming edge (except the start node)
A node can have multiple outgoing edges (branch nodes such as if-else)
Must ultimately connect to the answer node or other output nodes
sourceType and targetType must match the actual node type

Position coordinate layout

Recommended layout:

Starting X: 100
Node spacing X: 300-400
Y coordinate: keep on the same horizontal line (300) or adjust appropriately based on branches
Y coordinate of child nodes of branch nodes: ±150

Example layout:

Start(100,300) → LLM(400,300) → Code(700,300) → Answer(1000,300)

Branch layout:

                    → Branch1(1100,150)
If-Else(800,300) →
                    → Branch2(1100,450)

Workflow generation steps

1. Requirements analysis

Understand the user's business requirements
Determine the workflow type (chatflow/workflow/agent)
Identify the required node types
Plan the processing flow

2. Node design

Design the input variables of the start node
Plan the prompts for LLM nodes
Determine the code execution logic
Configure HTTP request parameters
Design the output format

3. Flow connection

Establish the logical relationships between nodes
Configure variable passing
Handle conditional branches
Ensure the flow is closed-loop

4. DSL generation

Generate YAML format that conforms to the specification
Configure unique node IDs
Set reasonable coordinate positions
Add necessary dependency plugins

5. Verification check

Check the correctness of the YAML format
Verify the completeness of variable references
Confirm the reasonableness of node connections
Check the completeness of required fields

Learning from real cases

Summary of learning based on 86+ real workflow cases:

Case 1: Image OCR recognition workflow

Requirements: Upload image → OCR recognition → Extract text

Node flow:

Start(file input) →
LLM(vision enabled, OCR) →
Answer(display result)

Key configuration:

Start node: type: file, allowed_file_types: [image]
LLM node: vision.enabled: true, vision.configs.variable_selector points to the file variable
System prompt: "Only output the text information recognized in the image"

Case 2: Text-to-HTML generation workflow

Requirements: Text description → Generate HTML code → Save as file

Node flow:

Start(text input) →
LLM(generate HTML) →
Parameter Extractor(extract HTML code) →
Tool(md_exporter, save file) →
Answer(return file URL)

Key configuration:

LLM prompt: "Generate an HTML program, output only the HTML code"
Parameter Extractor: Extract the HTML code block
Tool: Use the md_exporter plugin to save the file

Case 3: Data query visualization workflow

Requirements: User question → SQL query → Chart display

Node flow:

Start(query input) →
LLM(generate SQL) →
HTTP Request(database API) →
Code(format data) →
LLM(generate chart HTML) →
Answer(display chart)

Key configuration:

LLM1: Text-to-SQL generation
HTTP: Call the database query API
Code: Format JSON data into chart data format
LLM2: Generate ECharts/Chart.js HTML

Case 4: AI Agent workflow

Requirements: Complex task → Agent autonomous planning → Call tools → Return results

Node flow:

Start(task input) →
LLM(reasoning, with tools) →
Multiple Tool nodes(parallel) →
Variable Aggregator(collect results) →
LLM(summarize) →
Answer(final response)

Key configuration:

mode: agent-chat
LLM configuration: Enable multiple tools
Tool parallel execution: Multiple edges from LLM pointing to different Tools
Variable Aggregator: Aggregate the outputs of multiple Tools

Common prompt templates

Text-to-SQL

你是一个专业的 SQL 专家。根据用户的自然语言问题生成准确的 SQL 查询语句。

数据库schema:
{{#数据库结构#}}

用户问题: {{#sys.query#}}

要求:
1. 只输出 SQL 语句，不要有其他说明
2. 确保 SQL 语法正确
3. 使用合适的 JOIN 和 WHERE 条件

Data extraction

从以下文本中提取指定信息:

文本内容:
{{#输入文本#}}

提取要求:
- 提取所有日期
- 提取所有人名
- 提取所有金额

以 JSON 格式输出:
{
  "dates": [],
  "names": [],
  "amounts": []
}

HTML generation

根据用户需求生成完整的 HTML 页面。

需求: {{#用户需求#}}

要求:
1. 生成完整的 HTML 文档
2. 包含必要的 CSS 样式
3. 添加必要的 JavaScript 交互
4. 确保代码格式规范
5. 只输出 HTML 代码，不要 markdown 代码块标记

Notes

Rules that must be followed

Unique ID generation: Each node must have a unique ID (recommended to use a timestamp: 1747991890414)
Variable reference format: Must use {{#节点ID.变量名#}} format
Node type matching: The sourceType and targetType in edges must match the actual node types
Required fields: Required fields in YAML cannot be omitted
YAML format: Strictly follow YAML indentation rules (2 spaces)

Avoid common errors

❌ Variable reference error: {{节点ID.变量}} ✅ Correct format: {{#节点ID.变量#}}
❌ Missing node connection: Node is isolated and unconnected ✅ Ensure all nodes have connection relationships in edges
❌ Duplicate ID: Multiple nodes use the same ID ✅ Each node uses a unique ID
❌ Position overlap: Multiple nodes have the same coordinates ✅ Plan node positions reasonably to avoid overlap
❌ Missing dependency: A plugin is used but not declared in dependencies ✅ When using a plugin, the corresponding dependency must be added

Output format

The generated DSL file must be a complete YAML format that can be directly imported into Dify:

app:
  # App 配置

dependencies:
  # 依赖列表

kind: app
version: 0.3.0

workflow:
  conversation_variables: []
  environment_variables: []
  features:
    # 功能配置
  graph:
    edges:
      # 连接关系
    nodes:
      # 节点定义

Quality standards

Qualified standards (mandatory)

✅ YAML format is correct and can be parsed
✅ Includes complete app, dependencies, and workflow configuration
✅ Includes at least start and answer nodes
✅ Nodes have correct connection relationships
✅ Variable reference format is correct
✅ All required fields are complete

Excellent standards (recommended)

🌟 Prompt design is professional and meets business needs
🌟 Node layout is visually appealing and logic is clear
🌟 Includes appropriate error handling (if-else judgment)
🌟 Uses suitable plugins to enhance functionality
🌟 Code execution nodes are highly robust
🌟 Variable naming is semantic

Trigger keywords

Keywords that automatically trigger the dify-dsl-generator skill:

"Generate Dify workflow"
"Create Dify DSL"
"Dify YML file"
"Workflow configuration file"

Update log

v1.0.0 (2025-11-22)

✅ Initial version
✅ Based on learning from 86+ real cases
✅ Supports all major node types
✅ Complete DSL generation capability
✅ Intelligent node connection
✅ Standardized format output

Technical support

Reference resources:

Dify GitHub: https://github.com/langgenius/dify
DSL example repository: https://github.com/wwwzhouhui/dify-for-dsl
Dify official documentation: https://docs.dify.ai

About this skill

Core Features

Usage

Basic Usage

Detailed Usage

Dify DSL File Structure

1. App Configuration

2. Dependencies

3. Workflow

Node Type Details

Start Node

LLM Large Language Model Node

Code Execution Node

HTTP Request Node

If-Else Conditional Judgement Node

Tool Node

Answer direct reply node

Variable Aggregator node

Parameter Extractor node

Edges connection relationships

Position coordinate layout

Workflow generation steps

1. Requirements analysis

2. Node design

3. Flow connection

4. DSL generation

5. Verification check

Learning from real cases

Case 1: Image OCR recognition workflow

Case 2: Text-to-HTML generation workflow

Case 3: Data query visualization workflow

Case 4: AI Agent workflow

Common prompt templates

Text-to-SQL

Data extraction

HTML generation

Notes

Rules that must be followed

Avoid common errors

Output format

Quality standards

Qualified standards (mandatory)

Excellent standards (recommended)

Trigger keywords

Update log

v1.0.0 (2025-11-22)

Technical support