# ai-workflow-creator

> Generate CasareRPA workflows automatically from natural language prompts with automatic validation and self-healing. Creates workflow JSON using WorkflowBuilder, validates with test_workflow.py, and auto-fixes common errors. Use when: user asks to create workflow, automate task, build RPA flow, generate workflow from description.

- Author: CasareRPA Developer
- Repository: omerlefaruk/CasareRPA
- Version: 20260122083346
- Stars: 0
- Forks: 0
- Last Updated: 2026-02-06
- Source: https://github.com/omerlefaruk/CasareRPA
- Web: https://mule.run/skillshub/@@omerlefaruk/CasareRPA~ai-workflow-creator:20260122083346

---

---
name: ai-workflow-creator
description: Generate CasareRPA workflows automatically from natural language prompts with automatic validation and self-healing. Creates workflow JSON using WorkflowBuilder, validates with test_workflow.py, and auto-fixes common errors. Use when: user asks to create workflow, automate task, build RPA flow, generate workflow from description.
allowed-tools: Bash, Read, Write
---

# AI Workflow Creator

Generate complete CasareRPA workflow JSON files from natural language descriptions.

## Quick Start

When user asks to create a workflow:

1. **Parse requirements** - Extract automation steps from user's prompt
2. **Select nodes** - Choose appropriate nodes from [NODES.md](references/NODES.md)
3. **Build workflow** - Use WorkflowBuilder to create valid JSON
4. **Save file** - Output to `workflows/<name>.json`
5. **Execute first** - **CRITICAL: Run the actual workflow BEFORE validation to verify data flow**
6. **Verify output** - Check that the workflow produces the expected result (CSV, data, screenshots)
7. **Validate** - Run `test_workflow.py` and auto-fix if needed
8. **Self-heal** - Fix any issues and re-validate

**IMPORTANT:** Always execute the workflow to verify it produces the expected output. Validation only checks structure - execution checks behavior.

See [examples/](examples/) for complete workflow patterns.

## Critical Gotchas (READ BEFORE BUILDING)

### BrowserRunScriptNode Does NOT Store to Context Variables

**Problem:** `BrowserRunScriptNode` returns via output port `result`, but does NOT support `output_var` parameter.

**Solution:** Always use `SetVariableNode` to bridge the result to a context variable:

```python
# WRONG - output_var is ignored!
script_node = wb.add_node("BrowserRunScriptNode", config={
    "script": "result = [1, 2, 3]",
    "output_var": "my_data"  # DOES NOTHING!
})

# CORRECT - Use SetVariableNode to store
script_node = wb.add_node("BrowserRunScriptNode", config={
    "script": "result = [1, 2, 3]"
})
store_node = wb.add_node("SetVariableNode", config={
    "variable_name": "my_data"
})
wb.connect(script_node, "result", store_node, "value")
```

### LogNode Requires `data` Port

**Problem:** `LogNode` has a required `data` input port that causes validation errors if not connected.

**Solution:** Either provide data or skip LogNode entirely in generated workflows.

### Google Consent Dialog

**Problem:** Google shows a consent wall that blocks search results.

**Solution:** Handle consent in script or use DuckDuckGo as alternative.



## Browser Automation: Playwright Discovery (MANDATORY)

**For any browser automation workflow, FIRST use Playwright to discover actual selectors.**

### Step 1: Launch Discovery Session

```python
import asyncio
from playwright.async_api import async_playwright

async def discover_selectors(url: str):
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=False)
        page = await browser.new_page()
        await page.goto(url)

        # Pause for manual inspection
        print("Press Enter in terminal when ready to capture selectors...")
        input()

        # Capture available elements
        buttons = await page.query_selector_all("button")
        inputs = await page.query_selector_all("input")
        links = await page.query_selector_all("a")

        print(f"Found {len(buttons)} buttons, {len(inputs)} inputs, {len(links)} links")

        # Get specific element info
        for i, btn in enumerate(buttons[:5]):
            text = await btn.inner_text()
            print(f"Button {i}: '{text[:30]}...' - selector: button:nth-of-type({i+1})")

        await browser.close()

asyncio.run(discover_selectors("https://example.com"))
```

### Step 2: Test Selectors Work

```python
# Test before adding to workflow
element = await page.query_selector("#my-button")
if element:
    print("Selector works!")
else:
    print("Selector NOT found - try alternatives")
```

### Step 3: Update Workflow with Validated Selectors

Only use selectors that passed testing in the final workflow.

## Core Workflow Pattern

Every workflow needs:

```python
from casare_rpa.tools.workflow_builder import WorkflowBuilder

wb = WorkflowBuilder()
wb.workflow_data["metadata"] = {
    "name": "Workflow Name",
    "description": "Human-readable description",
    "version": "1.0.0",
    "author": "AI Workflow Creator",
    "tags": ["automation", "browser"],
}

# Position variables for clean layout
x, y = 0, 0
x_step = 250  # Horizontal spacing
y_step = 150  # Vertical spacing (if wrapping)

# Always start with StartNode
start_id = wb.add_node("StartNode", position=(x, y))

# Add automation nodes with proper spacing
x += x_step
node_id = wb.add_node("NodeType", config={"key": "value"}, position=(x, y))

# Connect: exec_out -> exec_in for flow
wb.connect(start_id, "exec_out", node_id, "exec_in")

# Save
wb.save("workflows/output.json")
```

## Visual Layout Guidelines

| Rule | Value | Reason |
|------|-------|--------|
| Horizontal step (x) | 250px | Prevents node overlap |
| Vertical step (y) | 150px | For multi-row layouts |
| Start position | (0, 0) | Top-left origin |
| Single row | All y=0 | Linear workflows |
| Multi-row | Increment y by y_step | Complex workflows |

## Common Node Patterns

| Task | Node | Config Example |
|------|------|----------------|
| Launch browser | `LaunchBrowserNode` | `{"headless": false, "browser": "chromium"}` |
| Navigate | `GoToURLNode` | `{"url": "https://example.com", "timeout": 30000}` |
| Click | `ClickElementNode` | `{"selector": "#button", "timeout": 5000}` |
| Type text | `TypeTextNode` | `{"selector": "#input", "text": "value"}` |
| Extract text | `ExtractTextNode` | `{"selector": ".content", "output_var": "result"}` |
| Wait | `WaitNode` | `{"duration_ms": 1000}` |
| Screenshot | `ScreenshotNode` | `{"file_path": "screenshots/capture.png"}` |

## Port Naming Rules

- **Flow ports**: `exec_in` / `exec_out` (always connect these for sequence)
- **Data ports**: node-specific (e.g., `selector`, `timeout`, `url`)

## Control Flow Patterns

### Loops
```python
# For loop over list
for_start = wb.add_node("ForLoopStartNode", config={"list_name": "items"})
for_end = wb.add_node("ForLoopEndNode")
wb.connect(for_start, "exec_out", "next_node", "exec_in")
wb.connect("next_node", "exec_out", for_end, "exec_in")
```

### Conditionals
```python
if_node = wb.add_node("IfNode", config={"condition": "${variable} > 0"})
merge_node = wb.add_node("MergeNode")
# Connect if branches to merge
```

### Error Handling
```python
try_node = wb.add_node("TryNode")
catch_node = wb.add_node("CatchNode", config={"error_var": "error"})
```

## Selector Strategies

| Situation | Selector |
|-----------|----------|
| ID | `#element-id` |
| Class | `.class-name` |
| Attribute | `[data-testid="submit"]` |
| Text contains | `text="Submit"` |
| Combined | `button.primary[type="submit"]` |

## Variable Usage

```python
# Set variable
set_var = wb.add_node("SetVariableNode", config={
    "variable_name": "my_var",
    "value": "static value"
})

# Get/use variable
get_var = wb.add_node("GetVariableNode", config={
    "variable_name": "my_var",
    "output_var": "result"
})

# In conditions, use ${variable_name} syntax
if_node = wb.add_node("IfNode", config={
    "condition": "${my_var} == 'expected'"
})
```

## Validation Checklist

Before saving, verify:
- [ ] Every workflow has `StartNode` (or starts execution)
- [ ] All exec connections form a valid path
- [ ] No circular dependencies (use loop nodes instead)
- [ ] All node_types exist in NODE_REGISTRY
- [ ] Selector strings are properly escaped
- [ ] Positions form a logical left-to-right flow

## Script Generation

For `BrowserRunScriptNode` or custom Python:

```python
wb.inject_script(node_id, """
# Multi-line script with automatic JSON escaping
result = await page.evaluate("document.title")
""")
```

## Testing & Self-Healing (MANDATORY)

**Every generated workflow MUST be tested and validated.**

### Step 1: Run Validation (Always)

After saving, immediately validate:

```bash
python .claude/skills/ai-workflow-creator/scripts/test_workflow.py workflows/output.json
```

### Step 2: Auto-Fix if Issues Found

If validation fails, run with auto-fix:

```bash
python .claude/skills/ai-workflow-creator/scripts/test_workflow.py workflows/output.json --fix
```

This automatically fixes:
- Duplicate node IDs
- Orphaned connections
- Self-connections
- Missing metadata

### Step 3: Background Execution Test

For browser workflows, run a background execution test:

```bash
# Run in background with timeout
python scripts/validate_test_workflows.py
```

### Step 4: Self-Healing Loop

If errors persist after auto-fix:

1. **Read the error report** - Note specific error codes
2. **Apply targeted fix** based on error code:
   - `ORPHANED_CONNECTION` → Remove connection or add missing node
   - `CIRCULAR_DEPENDENCY` → Use loop nodes instead of circular connections
   - `UNKNOWN_NODE_TYPE` → Check NODES.md for valid node name
   - `INVALID_PORT_NAME` → Use snake_case (exec_in, exec_out, etc.)
3. **Re-validate** until all tests pass

### Error Code Reference

| Code | Fix |
|------|-----|
| `ORPHANED_CONNECTION` | Remove connection or add missing node |
| `SELF_CONNECTION` | Remove self-connection |
| `DUPLICATE_NODE_ID` | Regenerate unique node IDs |
| `CIRCULAR_DEPENDENCY` | Use loop nodes (ForLoop/WhileLoop) |
| `UNKNOWN_NODE_TYPE` | Use valid type from NODES.md |
| `INVALID_PORT_NAME` | Use snake_case for ports |
| `UNREACHABLE_NODES` | Connect to main flow or remove |
| `EMPTY_WORKFLOW` | Add StartNode + action node |

## References

- [Complete node catalog](references/NODES.md) - All available nodes with configs
- [Workflow examples](examples/WORKFLOW_EXAMPLES.md) - Common patterns
- [Schema reference](references/SCHEMA.md) - WorkflowAISchema structure

## Output Format

Always save to `workflows/<descriptive-name>.json` with:
- Descriptive filename (kebab-case)
- Proper metadata
- Logical node positioning (left-to-right flow)
- Helpful comments in config where complex