In this post, we are officially graduating from basic AI prompting to a much robust level. Here, we’ll look at Google AI Studio—the web playground where advance users build and tweak prompts.
What Is Google AI Studio?
Think of Google AI Studio as the bridge between the consumer chat interface and raw Python code where you get raw access to the underlying models. You can select your model engine, test parameters, write your core instructions, and—crucially—generate the API keys required to connect software applications to Gemini.
1. AI Studio Playground
When you log into aistudio.google.com with your standard Google account, you’ll instantly see controls that don’t exist in consumer chatbots. Understanding these visual levers now makes writing the actual Python code later effortless.
Let’s get into the most powerful tool in the developer console of Google AI Studio i.e., System Instructions.
System Instructions
Located at the right side panel, this is where you permanently hardwire the model’s persona. Unlike standard chat prompts where you say “Act as an accountant,” System Instructions sit outside the normal conversation flow. The model never forgets them, ensuring it strictly sticks to your rules, tone, and forbidden phrases from the first turn to the last. When using a consumer chatbot, you waste a lot of space telling the model what not to do (“Don’t add intro text,” “Don’t hallucinate,” “Format it like this”). If the chat goes on too long, the model often forgets those rules.
System Instructions solve this. They act as permanent, global guardrails hardcoded into the model’s core architecture for that session. Today, we will build a specialized Data Cleaning & JSON Extraction Assistant completely inside the AI Studio UI to show you how this works.
When we use System Instructions:
- The model’s persona remains perfectly stable over long interactions.
- It strictly respects formatting constraints (like forcing output to be raw data only).
- It strips out conversational pleasantries, making it ideal for software pipelines.
Thinking Level
Thinking level in Google AI Studio is a setting that controls how much internal cognitive effort or reasoning a Gemini model applies before answering a prompt. Following are different thinking levels available:
- Low / Minimal: Best for high-throughput, latency-critical, or simple tasks like basic chat, data extraction, or content filtering where speed and cost matter most.
- Medium: Recommended default balance for everyday use cases, standard coding, and general multi-turn logic.
- High: Best for advanced mathematics, complex software architecture, or deep problem-solving that requires intense multi-step reasoning.
Now, let’s put these theories into action!
2. The Data Processing Agent
Here, we’ll build a custom agent designed to take messy, unformatted business or transaction logs and turn them into clean, structured data structures.
Step 1: Set the Foundation
Open Google AI Studio and click Create New Prompt (if not already in a fresh interface). In the right-hand configuration panel, make sure your model is set to Gemini 3.8 Flash (latest at the time of writing) and select the Thinking Level to Low (best for data extraction).
Step 2: Inject the System Instructions
Locate the System Instructions text box (usually at the right side of the prompt area). Paste the following master configuration:
You are a specialized, programmatic data-cleaning engine. Your sole function is to ingest messy text logs or transaction descriptions and convert them into structured data formats.
CRITICAL OPERATIONAL RULES:
1. ONLY output valid, parsable data blocks. Do not wrap data in conversational text.
2. NEVER include introductory or concluding phrases (e.g., do NOT say "Sure, here is your data").
3. If an input text is missing critical fields or is completely illegible, return an error field within the structured output: {"error": "Malformed input"}.
4. Normalize all dates to YYYY-MM-DD format.
5. Standardize all currency values to a float number with two decimal places.
6. Categorize each transaction into a definite accounting category.Step 3: Test with Messy Input
Now, move to the main user input box (at the lower side of interface). Let’s feed it a realistic, unstructured nightmare of a message that someone might copy-paste out of an email or an old chat history:
“Hey, just an update on expenses. Paid the software license to AWS on 23rd June 2026, ran us about Rs. 5,000. Also, grabbed lunch with the client at Pearl Continental yesterday (June 22) which was 6,500 PKR. Oh and someone lost the receipt for office supplies bought on May 12, I think it was around 4,000 rupees.”
Click Run at the bottom.
The Result: Pure, Predictable Output
Because of the low Thinking Level and the strict system instructions, Gemini doesn’t offer a chatty reply. It strips out the narrative and returns a perfectly structured data layout:
JSON
[
{
"date": "2026-06-23",
"description": "Software license - AWS",
"amount": 5000.00,
"currency": "PKR",
"category": "Software & Subscriptions"
},
{
"date": "2026-06-22",
"description": "Lunch with client at Pearl Continental",
"amount": 6500.00,
"currency": "PKR",
"category": "Meals & Entertainment"
},
{
"date": "2026-05-12",
"description": "Office supplies",
"amount": 4000.00,
"currency": "PKR",
"category": "Office Supplies"
}
]The information in that perfectly structured form is essential in order to feed the same into any system such as an accounting software.
3. The “Get Code” Button
Look at the top right corner of Google AI Studio interface. There is a magical button labeled Get Code.
Clicking it reveals something incredible: Google automatically converts the prompt, the system instructions, the thinking level settings, and the model choice you just configured into ready-to-run code in multiple languages.
If you click the Python tab, you’ll see exactly how the system_instruction parameter is fed into the initialization function. Cool!