Domain-encoded system prompts: turning 15 years of accounting forensics into a live Claude diagnostic tool

I built a structured diagnostic engine that ingests raw ADP Workforce Now payroll exports and NetSuite GL data, routes them through a domain-encoded Claude system prompt, and returns a JSON-structured root cause analysis — including the specific fix — for recurring payroll-GL discrepancies that human reviewers consistently miss.

The demo runs live against the Anthropic Messages API. The input is a real payroll pattern: employer burden accounts (CPP, EI, CNESST, RQAP) posting at exactly 0.500x expected in every quarter-end month for three years. The system prompt carries the diagnostic heuristics — check ratio before amount, look for N-period calendar patterns, separate gross wages from employer burden — that normally live inside a practitioner’s head. Claude doesn’t discover these rules. They’re injected. The output is a structured JSON object with root_cause, confidence_pct, adp_fix_steps[], correcting_je_description, and cfo_memo.