How I Built “Sunny,” a Scam-Savvy Chat Friend for Older Adults

Why Sunny?

Loneliness and digital fraud are both rising—especially for older adults. I wanted a tool that could be a friend first, with the emotional warmth and patience people deserve, and also a clear-headed co-pilot when something sketchy pops up on the phone or in the inbox. That’s how Sunny was born: a cheerful, plain-spoken companion that keeps conversation easy and can switch into scam-check mode on command.

Sunny is available for public use and feedback. On this page I’ll share how I built it, what’s under the hood, and lessons learned.

Design Goals

1. Friend first. Warm, upbeat tone, short replies by default, no judgment.
2. Scam-savvy. Structured questions for suspicious messages and calls; quick “Likely Scam / Needs More Info / Probably Legit” guidance with reasons.
3. Simple UX. No tech jargon. Every step is explained as if you’re helping a grandparent.
4. Safety & Boundaries. Never asks for passwords, SSNs, or bank details. Always reminds users to call official numbers on the back of their card or the organization’s real website.
5. Voice-friendly. Conversational pacing, polite turn-taking, and summary recaps that work well for speech.

What I Built It With

• OpenAI Custom GPT Builder. I used the built-in Instructions, Knowledge, and Conversation Starters to define Sunny’s personality and behaviors.
• Curated Knowledge. I uploaded short, plain-language references:
  • Common phone/email/text scam patterns
  • A step-by-step “scam triage” checklist
  • A conversational style guide (greeting patterns, gentle redirecting, recap templates)
  • A one-page disclaimer about limits and when to escalate to a real person
• Policies & Guardrails. I wrote explicit do’s/don’ts:
  • Do: Ask clarifying questions, suggest calling verified numbers, encourage involving a trusted family member.
  • Don’t: Give legal/financial/tax advice, diagnose medical issues, or collect sensitive information.
• Optional Voice Use. While Sunny runs inside ChatGPT today (tap the mic to talk), I also outlined how we’d expand to a real-time speech app later.

Building the Personality

I treated Sunny like a real teammate and wrote a style sheet:

• Tone: Cheerful, patient, polite.
• Length: Keep it short unless the user wants more.
• Pacing: Offer “Want me to keep going?” after key points.
• Empathy Rules: Acknowledge feelings first (“That sounded stressful.”), then help.
• Plain-Speak: Avoid acronyms; explain terms in one line.

Example openers Sunny uses:

• “Hi there, I’m Sunny. Want to chat or check a suspicious message together?”
• “We can take this one step at a time—no rush.”

Example closers:

• “If you want, I can list the next 2–3 steps.”
• “Would you like me to help you draft a short reply, or would you rather ignore and block?”

The Scam-Check Playbook (Built Into Sunny)

Sunny follows a repeatable, transparent flow:

1. Snapshot the situation. “What did they say? Did they ask for money, codes, or logins?”
2. Source check. “Where did this come from? Unknown number? Odd email domain?”
3. Pressure test. “Did they threaten legal action, account closure, or say ‘urgent’?”
4. Verification path. “Let’s independently call the company using the number on your card or their official website.”
5. Decision & next step. Sunny labels the risk level and gives 2–3 clear actions (block/report, ignore, or safe follow-up).

I wrote these as short, numbered prompts so Sunny delivers consistent, calm guidance every time.

Content Sunny Won’t Generate

Sunny is not a lawyer, doctor, or financial advisor. I set explicit refusals with friendly redirects:

• No legal/financial/tax decisions. Sunny explains why and suggests a qualified professional.
• No personal data intake. Sunny avoids collecting sensitive info and reminds users to safeguard codes and passwords.
• No technical troubleshooting that risks security. If it seems risky (e.g., unknown links), Sunny says so and proposes safer alternatives.

Testing With Real-World Scenarios

Before sharing Sunny publicly, I ran 30+ scenarios, including:

• “Harris County Sheriff” spoof calls. Sunny spotted number spoofing and recommended calling the real non-emergency line found on the official website.
• Bank account lockout texts. Sunny flagged generic greetings, mismatched URLs, and the urgency trap; guided users to log in via the bank’s app—never the text link.
• “Grandchild in trouble” calls. Sunny coached a calm callback plan using a known family number and suggested a shared family password for future verification.
• Charity appeals. Sunny checked IRS nonprofit status (conceptually) and suggested donating directly through the official site.

I kept iterating the prompts until Sunny’s answers were consistent, fast, and kind.

What Surprised Me

• Empathy matters more than features. Users opened up when Sunny acknowledged stress and offered small choices (“Want a short checklist or a quick call script?”).
• Short recaps reduce confusion. After a long explanation, Sunny gives a two-line summary—huge for clarity.
• Scripts help. Many people want exact words. Sunny provides simple scripts for blocking/reporting or checking with a company.

Privacy & Safety

• No storage of sensitive data. Sunny avoids requesting it.
• Escalation culture. When in doubt, Sunny recommends calling a verified number (from a bill, card, or official site), or looping in a trusted family member.
• Disclaimer. Sunny is informational support, not a substitute for professional advice or emergency services.

How to Use Sunny (Simple Instructions)

1. Open Sunny inside ChatGPT.
2. Talk or type. Start with “Sunny, can you check this text?” or just say “Hi.”
3. Show the details (safely). You can describe the message, paraphrase it, or paste text—skip personal numbers, codes, or passwords.
4. Follow the 2–3 next steps Sunny suggests. Sunny can also draft a polite reply or show how to block/report.
5. Use Sunny for companionship too. Sunny is always up for a friendly chat.

If You’re Building Something Similar

Here’s a quick checklist you can adapt:

• Define personality rules (tone, length, empathy) before anything else.
• Write a triage playbook as numbered steps the model can follow.
• Add hard guardrails: refusal patterns + safe redirects.
• Prepare micro-templates: short scripts for blocking, reporting, and verifying.
• Test against real scams (redact sensitive info) and iterate until answers are consistent and calm.
• Keep a simple disclaimer visible.

Closing Thoughts

Sunny isn’t just a scam shield. It’s a friendly presence that helps people feel seen, listened to, and in control—and it happens to be really good at cooling off high-pressure scams.

If you’d like to try Sunny, give feedback, or embed a similar helper for your organization, I’d love to connect.

Contact BoostMyAI: boost@boostmyai.com | 713-545-2882
Consulting & Custom Builds: Need a tailored companion or fraud-aware assistant for your community, property, or customer base? Let’s talk.