Georgia State University — J. Mack Robinson College of Business PATH — Pathways for AI Training & Hiring CIS 4394 Agentic AI  ·  Fall 2026  ·  Dr. Xinyu Fu
01 · Foundations

Anatomy of a prompt.

A prompt is everything the model sees before it responds — not just your question. Every well-built prompt has up to five parts, and each part is a knob you can turn (Boonstra, Google/Kaggle Prompt Engineering whitepaper, 2024–25).

Interactive

Click each part — see it in a real prompt

This is a real business prompt for routing support tickets. Click the five parts on the left and watch them light up.

You are a support-ticket triage assistant for Acme Software. Route tickets accurately; when unsure, choose HUMAN_REVIEW. Never invent order numbers. Context: Acme sells a project-management app. Billing handles payments and refunds; Technical handles bugs and access; Sales handles upgrades and demos. Example — ticket: "I was charged twice this month" output: {"category": "BILLING", "urgent": false} Return ONLY JSON: {"category": "BILLING"|"TECHNICAL"|"SALES"|"HUMAN_REVIEW", "urgent": true|false} Ticket: "The app deleted my project and I present to a client in an hour!!"
Click a part on the left.
Why it matters: when an output is wrong, don’t rewrite the whole prompt — diagnose which part failed. Wrong format? Fix the output spec. Wrong facts? Fix the context. Wrong behavior? Fix the instruction. That’s the skill Exercise A trains.
Two kinds of prompt

System prompt vs user prompt

System prompt

The employee handbook

Persistent. Set once by the developer; wraps every conversation. Role, scope, rules, refusals, output format. Users never see it — but it governs everything. (Week 2 flashback: you read Codex’s.)

User prompt

Today’s ticket

Per-turn. The specific question or task right now. Changes every message; interpreted under the system prompt’s rules.

Without a system prompt
User: Can I get a refund after 60 days? Model: Absolutely! Acme offers a 90-day money-back guarantee… ← invented policy

Not hypothetical: a tribunal made Air Canada honor a refund policy its chatbot invented — the airline argued the bot was “responsible for its own actions,” and lost. Ars Technica, Feb 2024 ↗

With a system prompt
SYSTEM: You are Acme support. Answer ONLY from the policy text provided. If the answer is not in the policy, say so and escalate. Never invent prices or policies. User: Can I get a refund after 60 days? Model: Our policy covers refunds within 30 days. I've flagged your case for a human agent to review options.
The system prompt is a policy guardrail. It’s also where Week 3 meets Week 2: prompt-level rules are the soft layer of guardrails; runtime enforcement (like Codex’s exec_policy) is the hard layer. Real systems need both — and the $1-Tahoe case (page 04) shows what happens with neither.
← BackWeek 3 home