From Prompt to Product: Build an AI Shopping Assistant in Minutes
A Hands-On Workshop with Google Gemini & AI Studio
Google Build With AI Event Series
Prasiddha Bista
Hello 👋 I’m Pras
Senior Site Reliability Engineer @Versent
Google Developer Expert (GDE) for GCP
Agenda
01
The Shopping Challenge
Understanding how AI can help you shop for products you love, saving you time and money
02
Prompt Engineering Essentials
Crafting inputs: The task, response format, and handling edge cases.
03
Mastering Google AI Studio
Navigating system instructions, grounding with search, and model settings.
04
The Hands-On Build
Building and deploying your own AI shopping assistant in real-time.
How Do You Shop for Something You've Never Bought Before?
We've all been there: you need a specific product, you don't know the market, and you end up spending 45 minutes comparing tabs, second-guessing prices, and still not sure if you're getting a good deal.
Hours lost comparing prices across multiple retailer websites
No single source of truth for "best value near me"
Decision fatigue from information overload
The insight: AI doesn't just know about products — when grounded with live search, it can shop for you, in seconds.
What Is Prompt Engineering?
A prompt is just instructions — the model follows them literally
Prompt engineering is the practice of crafting inputs to a language model to reliably produce the outputs you want. Think of it as writing a job description for an AI employee.
What to do
The task or goal
"Find prices for this product"
How to respond
Format, tone, length
"Return a markdown table"
Edge cases
Handling ambiguity or failure
"If no local results, show online options"
Why iteration matters: The same underlying model can produce wildly different outputs depending on how well the prompt is written. Today you will see this firsthand — three prompts, three dramatically different results.
Google AI Studio — Your Playground
No code needed. Just ideas.
Google AI Studio is a browser-based interface for building and testing Gemini-powered applications — no coding required.
System Instructions panel
Define your AI's role and behaviour before any conversation starts.
Grounding with Google Search
Connect Gemini to live web data so it retrieves real-time prices instead of relying on stale training data.
Shareable links
Deploy your assistant as a public URL with one click, no infrastructure needed.
Model settings
Adjust temperature, output length, and safety filters to fine-tune responses.
The Mental Model
System Instructions
Set the persona and rules. Written by you, the builder, to guide the AI's overall behaviour.
User Messages
The actual queries or prompts typed by the end-user during the conversation.
Keeping these separate is the foundation of building reliable AI apps.
What You're Building Today — Meet ShopBot
From a one-liner to a production-ready assistant in three steps
By the end of this workshop, your ShopBot will:
Search for live, real-time prices using Google Search grounding
Present results in a clean comparison table with store name, price, availability, and a buy link
Provide location-aware results with Google Maps directions
ShopBot Pick — a reasoned recommendation so users don't have to think
THE FINAL OUTPUT LOOKS LIKE THIS:
#
Store
Price
Availability
Link
1
Chemist Warehouse
$4.99
In stock
Buy Now
2
Woolworths
$5.50
Online
Buy Now
ShopBot's Pick
Chemist Warehouse offers the best value at $4.99 with in-store availability nearby.
The Prompting Progression — Three Steps, One Goal
You'll go from vague to polished by changing only the prompt
Prompt 1
WHAT CHANGES
Tell the model what it is
WHAT YOU GAIN
A baseline response — helpful but unstructured
Prompt 2
WHAT CHANGES
Tell the model what to do and how to format it + enable Google Search
WHAT YOU GAIN
Structured results with live prices and source links
Prompt 3
WHAT CHANGES
Give it a persona, a strict output format, and guardrails for edge cases
WHAT YOU GAIN
A polished, production-ready assistant with consistent behaviour
The underlying Gemini model is identical across all three prompts. The prompt is the product.
Prompt 1 — Give the Model a Role
Start simple: one sentence is enough to get going
System instruction
"You are a shopping assistant. Help users find the best prices for products they are looking for."
User message
"I am looking for a protein bar — Musashi Cookies and Cream (45g protein content)."
What you'll observe
The model is helpful and knowledgeable about the product
But results are vague — no real prices, no store names, no links
The format is a wall of text, hard to scan
The lesson
A role alone is not enough. The model needs to know how to respond, not just what it is.
Prompt 2 — Add Structure and Live Search
Enable Google Search grounding + explicit formatting instructions
Key additions to system instruction
"Use Google Search to find current, real-time prices and availability"
"Include store name, price, and a link to the product"
"Present results in a clear list format"
What changes
Results are now grounded in live web data — actual prices from actual retailers
The response is structured and scannable
Source links appear so users can verify and purchase directly
Enable Grounding with Google Search in the AI Studio model settings panel — this single toggle is the difference between stale training data and live, accurate prices.
Prompt 3 — Full ShopBot Persona
Persona + format + guardrails = a production-ready assistant
The final prompt adds three critical layers on top of Prompt 2 to create a complete experience:
Persona
ShopBot has a name, a clear purpose, and a consistent tone — "helpful, concise, and practical."
Structured Format
Every response follows the exact same template. No surprises for the user.
Product header
Comparison table
Edge Cases
Explicit instructions for handling ambiguity and missing data.
Category queries:
Asks clarifying questions first
No local results:
Flags clearly, shows online options
Deploy in one click
Hit Share → Create link in AI Studio to generate a public URL. No code. No servers. Done.
Share Link
Responsible AI — Trust, But Verify
Grounding is the mitigation; source links are the proof
Even the best AI shopping assistant can make mistakes. Here's what to watch for and how the workshop design addresses each risk:
Risk
Why it happens
How ShopBot mitigates it
Hallucinated prices
Model generates plausible-sounding but fabricated data
Google Search grounding anchors every price to a live source
Outdated information
Training data has a knowledge cutoff
Grounding fetches real-time web results at query time
Missing local stock
Search results may not reflect in-store availability
ShopBot flags uncertainty and provides online alternatives
Unverified links
URLs may change or expire
Always check the source link before purchasing
The golden rule: Never trust a price without a source link. ShopBot is built to always show its work.
https://bit.ly/4srnWMM
Workshop
What You Built Today — And What's Next
From zero to a deployed AI assistant in one workshop
What you accomplished
Wrote and iterated on three progressively better prompts
Learned the three levers of prompt engineering: task, format, and edge cases
Enabled real-time Google Search grounding for live, accurate data
Deployed a publicly accessible ShopBot with a shareable URL — no code required
Where to go from here
Swap the shopping domain for any use case: travel planning, recipe finder, local services
Explore Vertex AI for enterprise-grade deployments with access controls and audit logs
Experiment with temperature and safety settings in AI Studio to fine-tune behaviour
Even after GCP credits expire, Google AI Studio remains free to use
The prompt is the product. Keep iterating.
Thank you!
Do you have any questions?