legacy-system-automation

name: legacy-system-automation description: "Automate legacy systems via GUI automation, screen scraping, and RPA patterns" triggers: ["legacy", "GUI", "desktop", "screen", "RPA", "automation", "mainframe"] version: "1.0.0"

Legacy System Automation Skill

Overview

This skill teaches the Digital FTE how to automate legacy systems that lack modern APIs, using Computer Use (GUI automation), screen scraping, and RPA patterns.

Prerequisites

• Claude Agent SDK with Computer Use enabled
• Agent Sandbox (gVisor) for secure execution
• Screen resolution: 1920x1080 minimum

Core Procedures

1. System Discovery

Before automating any legacy system:

1. Identify the application: Name, version, access method (RDP, Citrix, local)
2. Map the workflow: Document each screen, field, and navigation path
3. Identify entry points: How to launch the application
4. Document expected states: What each screen should look like when ready

2. Screen Recognition Pattern

Use visual verification at each step:

1. Take screenshot
2. Identify UI elements (buttons, fields, menus)
3. Verify expected state matches actual state
4. If mismatch → retry screenshot, then flag for human review

3. Data Entry Procedure

For each field to populate:

1. Navigate: Click/tab to the correct field
2. Verify: Confirm cursor is in the right place (screenshot)
3. Enter: Type the data
4. Confirm: Verify data appears correctly (screenshot)
5. Proceed: Tab/click to next field or submit

4. Error Handling

5. Checkpoint Strategy

Save state after each major step:

{
  "step": "data_entry_complete",
  "timestamp": "ISO 8601",
  "screenshot_path": "/artifacts/screenshot_step3.png",
  "data_entered": {"field1": "value1", "field2": "value2"},
  "next_step": "submit_form"
}

6. Verification Protocol

After completing automation:

1. Take final screenshot
2. Compare against expected completion state
3. Extract confirmation number/message if present
4. Log success or failure with evidence

Security Constraints

• NEVER store credentials in screenshots or logs
• NEVER automate systems containing PII without sandbox isolation
• ALWAYS use Agent Sandbox (gVisor) for execution
• ALWAYS mask sensitive fields in screenshots before storing

Common Patterns

Pattern 1: Form Filling

Launch App → Navigate to Form → Fill Fields → Submit → Confirm

Pattern 2: Data Extraction

Launch App → Navigate to Report → Screenshot → Parse Data → Export

Pattern 3: Multi-Step Workflow

Step 1 → Verify → Step 2 → Verify → ... → Final Step → Confirm

Quality Metrics

• Accuracy: 99%+ field population accuracy
• Speed: Complete workflow within 2x human operator time
• Reliability: <1% failure rate requiring human intervention
• Evidence: Screenshot at every major step

Fallback Protocol

If automation fails:

1. Capture final screenshot with error state
2. Log all steps completed and where failure occurred
3. Queue task for human operator with full context
4. Notify via Slack MCP with error summary