Global Nerdy

It’s been about a week since Gemini 3 has come out, and I’ve been hearing some really good things about it. So I’m trying out Google’s “try Gemini 3 Pro for free for a month” offer and putting it to as much use as the account will allow, and I’m posting my findings here.

My approach to Gemini 3, after a week of heavy noodling with it

Gemini 3 feels fundamentally different from earlier text-based AI models, and I wouldn’t be surprised if its competitors eventually borrow at least a couple of tricks from it.

It’s designed to be an eater of large, not necessarily organized inputs such as text, images, video, audio, code, and documents simultaneously and natively. If you want to get the most of out it, I’ve found that a structured, almost engineering-like approach works better than simply treating it like just another chatbot.

Here are some things I and other Gemini 3 noodlers have noticed:

1. Get to the point and don’t sweat the etiquette

You’ve probably heard that adding “please” to your LLM prompts provides better results. It does work with many LLMs, and that’s they’ve been trained on examples of human communication, where politeness is often associated with higher-quality responses.

Gemini 3’s design as a “chaotic context eater that tries to find the signal in its input noise” means that it treats words like “please” as fluff. My experience is that unlike ChatGPT, it treats your prompt as a set of executable instructions rather than a chat.

“Chatting up” the model as if you were on a conversation over coffee doesn’t provide better results. What works best is providing it with the necessary context — that is, the background information needed to produce the desired result — and a clear goal stated clearly and without the fluff. “If you could please look at this file and tell me what you think” won’t work as well as “Analyze the attached PDF and list the critical errors that author made.”

Keep prompts short and precise. Google’s own guidance emphasizes that Gemini 3 responds best to direct instructions. Long prompts appear to divert the focus of the AI from the actual point and result in inconsistent. There is an area where you can “go long,” and that’s with the contextual information you provide prior to the actual instructions. I’ll talk about this later on in this article.

2. Gemini 3 answers are terse by default, but you can change that

Google is definitely trying to “zig” where other LLM vendors are trying to “zag” with the way Gemini 3 responds. It’s way more concise by default, and if you’re one of the people who was disappointed with the way OpenAI tuned ChatGPT 5.1 to be less chatty and obsequious, you’re going to be downright hurt by Gemini 3’s “ain’t nobody got time for chat” style of answer.

If you absolutely need a long, detailed narrative or a “chatty” persona, you’ve to got explicitly ask for it in your constraints. Otherwise, it will give you the shortest correct answer. And even then, I think Claude will give you more satisfying results in terms of conversational style.

Google has also stressed the Gemini 3 model’s tendency to respond in a neutral, efficient manner.  You want a friendly, funny, or conversational tone? Google says you have to ask for it. I’ve only played with this a little, and in my limited experience, it seems that the “friendly, funny, or conversational tone” feels like the tone used by someone at a service desk that’s trying to fake niceness in order to make the sale.

3. Anchor your behavioural constraints (or: “Expectations up front”)

If you need to spell out behavioral constraints to get the job done — such as “Be objective,” “Use a level of formality appropriate for an email to a lawyer,” “Don’t use external code libraries,” or “the code must be in Python 3.14+, and not in any earlier version” — do so either in the System Instruction or at the very top of your prompt. This ensures that these constraints set the direction of Gemini 3’s process before it starts crunching the data.

4. Name or index any info you give Gemini 3

Gemini 3 treats text, images, audio, and video as “equal-class inputs,” but now you have to eliminate any ambiguity that comes with giving it a big pile of information.

For example, if you upload three screenshots, a video,  and a PDF and then prompt with “look at this,” the model may struggle to understand which “this” you’re talking about. It “this” the first picture? The second one? The video? You need to be more specific.

Explicitly label your inputs in your prompt. Instead of vague references, say: “Use Image 1 (Funnel Dashboard) and Video 2 (Checkout Flow) to identify the drop-off point”. This forces the model to synthesize data across specific files.

Eevn better, provide the context info one upload at a time, giving a name to each as you upload it: “This is the Jenkins report,” “This is CSV of temperatures for October 2025.”

If you provide gemini 3 with data that has some index such as a timestamp or ID that identifies a specific piece of information within that data, use that index:

• With audio and video, refer to specific timestamps! For example, you can say “Analyze the user reaction in the video from 1:30 to 2:00.”
• With data that has built-in “coordinates”, use them!  For example, you can say “Analyze columns A-D in the CSV file.”

5. Need Gemini 3 to do deeper thinking? You have to make it.

For complex tasks, such as analyzing a legal contract or coding a simulation, you need to force Gemini 3’s model to slow down and “think” before it generates a final response. This prevents hallucination and ensures logical consistency.

• Explicit decomposition is your friend. Don’t just ask for the solution. Break big “thinky” tasks into smaller sub-tasks and have Gemini execute those. For example, ask it to “Parse the stated goal into distinct sub-tasks” first. By forcing it to create an outline or a plan, you improve the quality of the final output.

• Tell it to be its own critic. Instruct the model to critique its own work before finishing. Include a step in your instructions like: “Review your generated output against the user’s original constraints. Did I answer the user’s intent?”

• Make Gemini 3 do a little self-reflection. If the model needs to use a tool (like writing code or searching), ask it to explicitly state why it is using that tool and what data it expects to find before it executes the action.

Previous articles in this series

• Part 1: Context entropy and a huge context window
• Part 2: It’s multimodal!

It’s been a week since Gemini 3 has come out, and I’ve been hearing some really good things about it. So I’m trying out Google’s “try Gemini 3 Pro for free for a month” offer and putting it to as much use as the account will allow, and I’m posting my findings here.

What does it mean for an AI model to be multimodal?

Multimodal, in the AI sense of the word, means “capable of understanding and working with multiple types of information, such as text, images, audio, and video.” We humans are multimodal, and we’re increasingly expecting AIs to be multimodal too.

The payoff that comes from an AI model being multimodal is flexibility and naturalness. There are times when it’s better to provide a picture instead of trying to type a description of that picture, or give the model a recording of a discussion rather than transcribing it first.

Gemini 3 is natively multimodal

The marketing and developer relations teams behind AI models will claim that their particular product is  multimodal, but a number of them are only indirectly that way.

Many models are strictly language models (the second “L” in “LLM”). These models use other models to translate non-text information to text, including models that “see” an image and generate text to describe it, and then use that text as the input. This approach works, but as you might have guessed, a lot gets lost in the translation. The loss is even greater with video.

Gemini 3 is different, since it’s natively multimodal. That means that it processes text, images, audio, video, and code simultaneously as a single stream of information, and without translating non-text data into text first.

How to use Gemini 3’s multimodality

1. Coding multimodally

Because Gemini 3 processes all inputs together, it can reason across them in ways other models can’t. For example, when you upload a video, it can analyze the audio tone against the facial expressions in the video while cross-referencing a PDF transcript of that video.

This opens a lot of interesting possibilities, but I thought I’d go with a couple of ideas I want to try out when coding:

• On-brand user interfaces: I could upload a hand-drawn sketch or a screenshot of a website I liked and say: “Build a functional React front end that captures this ‘look,’ but uses my company’s brand colors, which are defined in this style guide PDF.”

• Multimedia debugging: When debugging with a text-first AI model, I’d simply give it the error log. But with a multimodal model like Gemini 3, I could provide additional info for more context, including:

  • The code
  • The raw server logs

  • A screen recording of the application that includes the moment the bug rears its ugly head, I could even ask: “Locate the specific line of code in File A that’s causing the visual glitch at 0:15 in the video.”

2. Get multimodal answers

Here’s one straight out of science fiction. Instead of asking for text answers, try asking for an answer that includes text, graphics, and interactivity.

For example, I gave it this prompt:

Don’t just explain mortgage rates. Code me an interactive loan calculator that lets me slide the interest rate and see the monthly cost change in real time.

It got to work, and in about half a minute, it generated this app…

…and yes, it worked!

3. Using multimodality and Gemini 3’s massive context window

In my previous article on Gemini 3, I said that it had a context window of 1 million tokens. When you combine that with multimodality, you get a chaos crushing machine. You don’t have to “clean” your data before hading it over to the model anymore!

• The haystack search: Upload an entire hour-long lecture video, three related textbooks, and your messy lecture notes, and then prompt it with: “Create a study guide that highlights the 5 concepts from the video that are NOT covered in the textbooks.”

• Archiving in multiple languages: Upload handwritten recipes or letters in different languages. Gemini can decipher the handwriting, translate it, and format it into a digital cookbook or archive. For a cross-cultural family like mine, this could come in really handy.

Multimodal tips

To get the best results…

• Name your inputs. Don’t say “look at the image.” Provide names for images when you upload them (“This is image A…”) and then refer to them when providing instructions (“Using image A…”).

• Give it the info first, instructions last. When taking advantage of that huge context window, provide Gemini with all the data you want it to use first, and put your instructions at the end, after the data. Use a bridging phrase like “Based on the info above…”

• With video, use timestamps. If there’s a part of the video that’s relevant to your instructions, refer to it by timestamp (e.g., “The trend visible at 1:45 in video B…”).

Previous articles in this series

• Part 1: Context entropy and a huge context window

It’s been about a week since Gemini 3 has come out, and I’ve been hearing some really good things about it. So I’m trying out Google’s “try Gemini 3 Pro for free for a month” offer and putting it to as much use as the account will allow, and I’m posting my findings here.

Gemini 3 and context entropy

Gemini 3, probably owing to its roots at Google, seems to handle “context entropy” better than ChatGPT or Claude. By “context entropy,” I mean “messy data,” with the disorder and chaos that you expect to find in notes and documents that you accumulate over time. Unless you’ve put in a lot of time, you probably haven’t put this information into much of a structured form or created some kind of “map” that explains how (or even if) the various parts of the information are related to each other.

Gemini 3 does a good job of taking chaotic information and finding the signal. I recently fed it a collection of…

• Job descriptions for positions that I’ve applied for
• Various versions of my resume, each one tuned for a specific job application
• Cover letters for each job application
• Screenshots of job application forms, where I had to answer -pre-screening questions posed by prospective employers
• Notes from interviews with recruiters and prospective employers
• Video recordings of the aformentioned interviews, which took place on Zoom, Teams, or Google Meet (with the approval of the other parties, of course)
• Slide presentations that I gave as part of the interview process
• Follow-up emails, texts, and other messages
• “Conversations” with ChatGPT and Claude through the job search process, from generating customized cover letters and resumes all the way to post-mortems

…and it’s produced some useful stuff, including a strategy document that I plan to use in my job search going forward. (More on that in a later post here and video on the Global Nerdy YouTube channel.)

Gemini 3’s huge context window

The other notable thing about Gemini 3 is its context window, which is an AI model’s working memory, or the maximum amount of information that it can work with in any given chat session. Gemini 3 Pro’s context window is a huge 1 million tokens, which is roughly equivalent to about 750,000 words (a token is a small chunk of language that lies somewhere between a character and a full word).

This big context window means that it’s possible to feed Gemini 3 with a LOT of data, which could be a big application codebase, a couple of textbooks, lots of notes or transcripts, or any other big pile of messy data that you’re trying to extract meaning from.

I’m going to be working heavily with Gemini 3 over the next few weeks, and I’ll continue to post my observations here, along with tips and tricks that I either find online or figure out. Watch this space!