Dify Workflow Builder

Expert skill for generating production-quality Dify DSL workflow files. Generates valid YAML configurations for Dify v1.12+ with precise node schemas derived from Pydantic source models.

Reading Strategy

Always read:

1. This file (SKILL.md) — process, patterns, quality standards
2. Node Index — base fields + pick which node files to load
3. Entry nodes + Output nodes — every workflow needs these

Read on demand (only the files relevant to your task):

• Node type files from references/nodes/ — only for node types used in the workflow
• Edge Types — when connecting nodes (sourceHandle rules)
• Variables — when using system/env/conversation variables
• Workflow Structure — for features, conversation_variables, environment_variables
• Node Positioning — for layout constants
• Enums — for exact enum values
• API Reference — for import/export API calls

Example workflows in assets/ — read the most relevant one as a template before generating.

Mode Selection

Key differences:

• workflow mode: No sys.query, no conversation memory, uses end node with explicit outputs
• advanced-chat mode: Has sys.query, sys.conversation_id, sys.dialogue_count, uses answer node for streaming response
• Only advanced-chat supports conversation_variables for session-persistent state

Workflow Generation Process (5 Steps)

Step 1: Define Type and Inputs

app:
  mode: workflow | advanced-chat
  name: "Workflow Name"
  description: "What this workflow does"
  icon: "🤖"
  icon_background: "#FFEAD5"
kind: app
version: "0.5.0"

Define start node variables based on user requirements:

- data:
    type: start
    title: Start
    variables:
      - type: text-input | paragraph | select | number | file | file-list
        variable: input_name
        label: "Human-readable label"
        required: true

Step 2: Design the Node Graph

Map the business logic to node types. Read the relevant files from references/nodes/:

• AI processing: ai.md — llm, agent, question-classifier, parameter-extractor
• Data: data.md — knowledge-retrieval, datasource, document-extractor
• Logic: logic.md — if-else, code, template-transform, http-request, tool, list-operator
• Flow control: flow.md — iteration, loop, human-input
• Variables: variables_nodes.md — variable-aggregator, assigner

Step 3: Generate Unique IDs

Every node needs a unique ID. Use the timestamp-based format:

node_id = str(int(time.time() * 1000))[-10:]  # e.g. "1732456789"

Step 4: Configure Edges

Connect nodes with edges. Read Edge Types for full sourceHandle rules.

- data:
    sourceType: llm
    targetType: end
    isInIteration: false
    isInLoop: false
  id: "source_id-target_id-sourceHandle-targetHandle"
  source: "source_node_id"
  sourceHandle: source        # "source" | "{case_id}" | "false" | "success-branch" | "fail-branch"
  target: "target_node_id"
  targetHandle: target        # always "target"
  type: custom

Step 5: Add Positioning

See Node Positioning for full layout constants.

Linear chain: each node at x += 300 (NODE_WIDTH_X_OFFSET), starting at {x: 80, y: 282}.

  position:
    x: 80
    y: 282
  width: 244
  height: 98  # varies by node type

Common Workflow Patterns

Pattern 1: Simple LLM Chain

start → llm → end/answer

See assets/simple_llm_workflow.yml or assets/simple_llm_chatflow.yml.

Pattern 2: RAG (Retrieval-Augmented Generation)

start → knowledge-retrieval → llm → answer

See assets/knowledge_rag_chatflow.yml.

Pattern 3: Conditional Branching

start → if-else → [branch A: llm-1] → variable-aggregator → end
               → [branch B: llm-2] ↗

See assets/conditional_workflow.yml.

Pattern 4: Error Handling

start → llm (error_strategy: fail-branch)
          → [success-branch] → end
          → [fail-branch] → variable-aggregator → end

See assets/error_handling_workflow.yml.

Pattern 5: Iteration

start → iteration [llm → code] → end

See assets/iteration_workflow.yml.

Pattern 6: Loop with Break

start → loop [code → if-else → loop-end] → end

See assets/loop_workflow.yml.

Pattern 7: Agent with Tools

start → agent(tools: [web-search, calculator]) → answer

See assets/agent_chatflow.yml.

Pattern 8: Human-in-the-Loop

start → llm → human-input → code → end

See assets/human_approval_workflow.yml.

Quality Standards

Required (MUST)

1. version: "0.5.0" — current DSL version
2. mode: workflow | advanced-chat — not agent-chat
3. Every node has unique id (string, timestamp-based)
4. Every edge references existing source and target node IDs
5. workflow mode uses end node, advanced-chat uses answer node
6. All variable_selector paths resolve to existing node outputs or system variables
7. sourceHandle matches node type pattern (see Edge Types)
8. error_strategy only uses "fail-branch" or "default-value" (NOT "abort" or "retry")
9. Container nodes (iteration, loop) have start_node_id pointing to internal start node
10. assigner (v2) has version: "2" field

Recommended (SHOULD)

1. Use descriptive title for each node
2. Enable retry_config for unreliable external calls (http-request, tool)
3. Use error_strategy: "fail-branch" for non-critical paths
4. Set is_parallel: true for iteration when items are independent
5. Keep parallel_nums ≤ 10 to avoid rate limits
6. Validate with scripts/validate_workflow.py before import