bpmn-generator

name	bpmn-generator
description	Generate BPMN 2.0 compliant XML files from natural language process descriptions OR from structured markdown business process documents. Use this skill when a user wants to create a BPMN workflow, convert a business process to BPMN XML, model a workflow diagram, or generate process definitions. Triggers on requests like "create a BPMN", "generate workflow XML", "model this process", "convert to BPMN 2.0", "create process diagram", "build workflow", or "convert this markdown to BPMN". Do NOT use for converting an existing BPMN XML file to Draw.io — use bpmn-to-drawio for that.
argument-hint	[<markdown-file>] or process description
allowed-tools	Read, Write, Glob, Grep, AskUserQuestion

BPMN 2.0 XML Generator

Overview

This skill transforms process descriptions into fully compliant BPMN 2.0 XML files. It operates in two modes:

Mode	Trigger	Workflow
Interactive	Natural language description, no file provided	Structured Q&A to gather requirements
Document Parsing	Markdown file path provided	Parse document structure, extract elements

The generated XML includes:

• Complete process definitions with all BPMN elements
• Proper namespace declarations for BPMN 2.0 compliance
• Diagram Interchange (DI) data for visual rendering
• Phase comments for PowerPoint generation compatibility
• Layouts compatible with Draw.io, Camunda, Flowable, and bpmn.io

---

MODE DETECTION

Automatic Mode Selection

Determine the operating mode based on user input:

IF user provides a markdown file path (.md):
    → Document Parsing Mode
ELSE IF user provides a natural language description:
    → Interactive Mode

Document Parsing Mode Indicators

• File path ending in .md
• "convert this document", "parse this file"
• "generate BPMN from [filename]"
• Markdown content pasted directly

Interactive Mode Indicators

• Brief process description without file
• "create a BPMN for...", "model a process that..."
• Questions about process design
• No structured document provided

Preview Mode

Both modes support an optional --preview flag:

When --preview is specified:

1. Generate the complete BPMN XML in memory
2. Validate structure (namespace, elements, flows)
3. Display summary:

   Preview: /bpmn-generator
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
   Process: Order Fulfillment
   Source: Interactive mode
   
   Structure Summary:
     Pools: 1
     Lanes: 3 (Sales, Operations, Shipping)
     Tasks: 8 (5 user, 3 service)
     Gateways: 2 (1 exclusive, 1 parallel)
     Events: 2 (1 start, 1 end)
   
   Validation: PASSED
   Output file: order-fulfillment.bpmn
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
   
   

4. Confirm with AskUserQuestion before writing anything — a single question, header: "Save", options Save the file (recommended, states the output path) and Discard (exits without writing). Wait for the response.
5. On Discard or a skipped question: exit without saving

---

PART 1: INTERACTIVE MODE

Use this mode when the user provides a natural language description without a structured document.

Interactive Question Framework

Purpose

Initial process descriptions are rarely sufficient for optimal BPMN generation. This mode uses a structured clarification process to gather complete requirements before generating XML.

Question Format

Every clarifying question is asked with the native AskUserQuestion tool — one question per call, 2–4 real options, the recommended one first and labelled (Recommended). Never hand-roll a lettered menu, and never add a "provide your own" or "skip" option: the harness supplies a free-text Other box and a Skip control on every question.

The canonical shape, the 22 ready-made question blocks, and the rules that govern them live in references/clarification-patterns.md (Question Format Template). Read it before asking the first question and instantiate its blocks rather than inventing new ones.

Auto-Accept Mode

Reached through the free-text Other box at any question — see Auto-Accept Mode Behavior in references/clarification-patterns.md, which owns the trigger, the recommended-answer rule, and the decisions-summary table rendered before XML generation.

Question Phases

Process questions in this specific order:

Phase 1: Process Scope (Questions 1-3)

• Process name and identifier
• Process trigger (start event type)
• Process completion states (end event types)

Phase 2: Participants (Questions 4-5)

• Single process vs. collaboration (multiple pools)
• Lanes/roles within pools

Phase 3: Activities (Questions 6-10)

• Main activities/tasks identification
• Task types for each activity
• Task descriptions/documentation (CRITICAL for PowerPoint generation)
• Task sequencing and dependencies
• Subprocess candidates

Phase 4: Flow Control (Questions 11-15)

• Decision points requiring gateways
• Gateway types (exclusive, parallel, inclusive, event-based)
• Default flows
• Loop/cycle detection

Phase 5: Events & Exceptions (Questions 16-19)

• Intermediate events (timer, message, signal)
• Boundary events on tasks
• Error handling approach
• Compensation requirements

Phase 6: Data & Integration (Questions 20-22)

• Data objects needed
• External system integrations
• Message flows (for collaborations)

Phase 7: Optimization Review (Question 23)

• Final review of proposed structure
• Opportunity for adjustments

Session Control

There is no session-command interpreter. The native UI already provides what the old help/status/back/skip/quit REPL simulated: Skip is the skip command, closing the question is quit, and status is redundant when the interview is visible in the transcript. back has no native equivalent — if the user asks to revisit an earlier question in free text, re-ask it and overwrite the recorded decision.

Adaptive Questioning

Skip questions that don't apply:

• Skip participant questions for simple single-pool processes
• Skip data questions if no data dependencies mentioned
• Skip error handling if process is straightforward
• Always ask critical questions: start event, main tasks, end events

---

PART 2: DOCUMENT PARSING MODE

Use this mode when the user provides a markdown file containing a structured business process document.

Document Analysis Steps

Step 1: Identify Document Structure

Analyze the markdown document for structural elements that map to BPMN constructs. Key patterns: H1 = process name, H2/H3 "Phase/Step" = phase comments, numbered lists = tasks, role tables = lanes, conditional language = gateways, "begins when" = start events, "completes when" = end events. See ../references/markdown-parsing-guide.md for the complete document structure mapping table.

Step 2: Extract Process Metadata

From the document, extract:

process_name: [from H1 or title]
process_id: [sanitized process_name, e.g., "SocialMediaCommunityManagement"]
description: [from executive summary or overview section]
version: [from document metadata if present]
roles: [list of all mentioned roles/actors]
phases: [ordered list of phase names from section headings]

Step 3: Map Roles to Lanes

Use the lane mapping configuration in ../templates/lane-mapping.yaml to assign colors. Read references/bpmn-elements.md for the complete role-to-lane color mapping table (Sales, Legal, Finance, IT, Implementation, Training, Customer Success, Support, Customer).

Step 4: Parse Phases and Tasks

Phase Detection Patterns

Match headings like "## Step 1:", "### Phase 2:", or "## 1.1 Section" using H2-H4 with step/phase/stage keywords or numbered sections. See ../references/markdown-parsing-guide.md for regex patterns.

Task Type Inference

Use the task type selection table in references/bpmn-elements.md to map markdown language to BPMN task types. Common shortcuts: "reviews/approves" = userTask, "system/API" = serviceTask, "sends/notifies" = sendTask, "waits for/receives" = receiveTask, "subprocess" = subProcess.

Step 5: Extract Documentation

Critical: Every task MUST have a <bpmn:documentation> element. Extract from:

1. Paragraph following task heading
2. Bullet points under task name
3. Table cell descriptions
4. "Process Description:" sections

Combine multiple sources into comprehensive documentation:

<bpmn:userTask id="Activity_ReviewTriage" name="Review and Triage">
    <bpmn:documentation>
        Community Manager and Social Team Lead manually review inbound interactions
        to assess and categorize them. Triage criteria includes: Topic/Intent,
        Location Relevance, Urgency, Risk Level, and Sentiment. Categories include
        location-specific issues, digital inquiries, brand questions, escalations,
        spam, and general engagement.
    </bpmn:documentation>
</bpmn:userTask>

---

PART 3: SHARED BPMN GENERATION

Both modes use the same BPMN generation rules.

BPMN Element Mapping

Read references/bpmn-elements.md (relative to this plugin's directory) for element type mappings and DI constants, including:

• Task Type Selection - keyword-to-BPMN-task-type mapping (userTask, serviceTask, sendTask, etc.)
• Gateway Selection - decision-pattern-to-gateway-type mapping (exclusive, parallel, inclusive, event-based)
• Event Selection - start and end event type mappings with XML elements

Phase Comments (CRITICAL for PowerPoint)

Always include phase comments to enable automatic phase detection for PowerPoint presentations:

<bpmn:process id="Process_Example" name="Example Process" isExecutable="true">

    <!-- Phase 1: Intake and Validation -->
    <bpmn:startEvent id="StartEvent_1" name="Request Received">
        ...
    </bpmn:startEvent>

    <!-- Phase 2: Processing -->
    <bpmn:serviceTask id="Activity_Process" name="Process Request">
        ...
    </bpmn:serviceTask>

    <!-- Phase 3: Fulfillment -->
    <bpmn:userTask id="Activity_Fulfill" name="Fulfill Request">
        ...
    </bpmn:userTask>

</bpmn:process>

XML Generation Rules

Required Structure

<?xml version="1.0" encoding="UTF-8"?>
<bpmn:definitions
    xmlns:bpmn="http://www.omg.org/spec/BPMN/20100524/MODEL"
    xmlns:bpmndi="http://www.omg.org/spec/BPMN/20100524/DI"
    xmlns:dc="http://www.omg.org/spec/DD/20100524/DC"
    xmlns:di="http://www.omg.org/spec/DD/20100524/DI"
    xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
    id="Definitions_[unique-id]"
    targetNamespace="http://bpmn.io/schema/bpmn"
    exporter="Claude BPMN Generator"
    exporterVersion="2.0">

    <!-- Process definition goes here -->

    <!-- Diagram interchange goes here -->

</bpmn:definitions>

ID Generation Rules

Read references/bpmn-elements.md for the complete ID pattern table (Process, StartEvent, EndEvent, Activity, Gateway, Lane, Flow patterns with examples).

Sequence Flow Rules

1. Every element (except start events) MUST have at least one incoming flow
2. Every element (except end events) MUST have at least one outgoing flow
3. Gateways splitting must eventually merge (except for end paths)
4. Conditional flows MUST have condition expressions:

<bpmn:sequenceFlow id="Flow_1" sourceRef="Gateway_1" targetRef="Task_2">
    <bpmn:conditionExpression xsi:type="bpmn:tFormalExpression">
        ${condition == true}
    </bpmn:conditionExpression>
</bpmn:sequenceFlow>

5. Default flows from gateways:

<bpmn:exclusiveGateway id="Gateway_1" default="Flow_default">
    ...
</bpmn:exclusiveGateway>
<bpmn:sequenceFlow id="Flow_default" sourceRef="Gateway_1" targetRef="Task_3"/>

Diagram Interchange Generation

Element Dimensions and Layout Constants

Read references/bpmn-elements.md for element dimension tables (width/height for events, tasks, gateways, subprocesses) and Draw.io-compatible layout constants (pool label width, lane offsets, element spacing).

Cross-Lane Edge Rule (CRITICAL)

Edges crossing lane boundaries MUST have parent="1" (root) with absolute mxPoint coordinates when targeting Draw.io conversion.

Validation Checklist

Before outputting XML, verify:

Structural Integrity

• Exactly one start event (or multiple for event subprocess)
• At least one end event
• All elements connected via sequence flows
• No orphaned elements
• All IDs unique within document

Flow Validity

• Start events have no incoming flows
• End events have no outgoing flows
• All other elements have both incoming and outgoing flows
• Parallel splits have matching parallel joins
• No infinite loops without exit condition

BPMN 2.0 Compliance

• All required namespaces declared
• All elements have required attributes (id, name where applicable)
• Conditional flows have condition expressions
• Default flows properly marked on gateways
• Event definitions properly nested

Documentation & Phases

• All tasks have <bpmn:documentation> elements
• Phase comments included for PowerPoint compatibility
• Documentation is comprehensive (not just task name repeated)

Diagram Interchange

• Every process element has corresponding BPMNShape
• Every sequence flow has corresponding BPMNEdge
• All shapes have valid Bounds (x, y, width, height)
• All edges have at least 2 waypoints
• No negative coordinates
• Elements don't overlap

---

OUTPUT FORMAT

For Interactive Mode

1. Decision Summary

Display a "Process Configuration Summary" with process name, ID, and a decisions table (# / Topic / Decision).

2. Process Description

Show a brief narrative of the process flow with a flow summary line (Start -> Task -> Gateway -> End).

3. BPMN XML File

Write the complete XML to a file named [process-name].bpmn in the current directory.

For Document Parsing Mode

1. Conversion Summary

Display source document, process name/ID, extracted structure counts (phases, roles/lanes, tasks by type, gateways, events), and any assumptions made during parsing.

2. BPMN XML File

Write complete XML to [process-name].bpmn

Common Output

Validation Confirmation

After generating XML, display validation results confirming: structural checks passed, flow validity passed, BPMN 2.0 compliance verified, diagram interchange complete, phase comments included, and the output filename.

---

Performance

Process Complexity	Elements	Expected Duration	Notes
Simple (linear flow)	5-10	1-3 minutes	Few gateways, single pool
Medium (branching)	10-30	3-8 minutes	Multiple gateways, 2-4 lanes
Complex (collaboration)	30-60	8-15 minutes	Multiple pools, message flows, subprocesses
Document parsing mode	varies	2-5 minutes	Faster than interactive; no Q&A overhead

Interactive mode duration is dominated by the Q&A clarification phases (user response time not included). XML generation and DI coordinate calculation add 10-30 seconds regardless of complexity. Document parsing mode is faster because it skips the interactive question framework and extracts structure directly from markdown.

Error Handling

• No file path and no process description provided: Ask the user for one before proceeding — do not guess a process or mode (see Mode Detection).
• Document Parsing Mode input doesn't match any known structural pattern (no headings, no numbered steps, no role table): fall back to Interactive Mode and tell the user why, rather than emitting a mostly-empty BPMN file.
• Validation Checklist failures (orphaned elements, missing flows, undocumented tasks) before writing output: fix the specific element or ask the user for the missing detail — never emit non-compliant XML.
• Preview Mode declined (n/no): exit without writing any file.
• Unrecognized session command during Interactive Mode Q&A: show the help message (see Session Commands) rather than silently treating it as an answer choice.
• Referenced support files missing (../references/*.md, ../templates/*): report which file is missing and proceed using the inline defaults documented in Part 3 rather than failing outright.

REFERENCES

For detailed specifications, see:

• ../references/bpmn-elements.md - Element type mappings, DI constants, ID patterns, and lane colors
• ../references/bpmn-elements-reference.md - Complete element catalog
• ../references/xml-namespaces.md - Namespace documentation
• ../references/clarification-patterns.md - Question templates (Interactive mode)
• ../references/markdown-parsing-guide.md - Document parsing patterns (Document mode)

For templates, see:

• ../templates/bpmn-skeleton.xml - Base structure
• ../templates/element-templates.xml - Element snippets
• ../templates/lane-mapping.yaml - Role to lane color mapping

For examples, see:

• ../examples/ - Complete working examples for both modes