compliance-check

Use whenever output to customers, prompts, or tool behavior might touch financial advice, regulation, or sensitive data. Enforces the spec's compliance rules — what is/isn't allowed.

SKILL.md

GitHub

Compliance Check skill

When to invoke

  • Reviewing any customer-facing text (WhatsApp/email/CRM)
  • Designing or modifying system prompts
  • Adding new tools that could produce financial advice
  • Reviewing escalation logic

Hard rules (non-negotiable, from spec)

  1. Never fabricate numbers (returns, balances, fees, tax rates)
  2. Never quote a specific regulation/section without citing source
  3. Never give a specific investment recommendation
  4. Never discuss a customer's case with a different customer
  5. Personalized financial advice requires explicit human licensee approval — Claude prepares draft, advisor approves

Escalation triggers (auto-route to human)

  • Regulatory question ("האם מותר לי...", "מה החוק אומר")
  • Customer crisis (anger, threat, complaint to regulator, "אני אתבע")
  • Compliance doubt (uncertain whether response is allowed)
  • Missing critical data (advisor name, customer_id, process stage)
  • Personal advice request ("מה הכי כדאי לי?", "תגיד לי איזה מסלול")
  • API failure that affects a customer's process

Review checklist for any prompt change

  • Does the prompt instruct to refuse fabrication?
  • Does it list escalation triggers explicitly?
  • Does it require Hebrew output for customer-facing channels?
  • Does it forbid quoting unverified figures?
  • Is there a clear "when in doubt, escalate" rule?