AI Live Assistance Google Sheets & Coding w/ Gemini Hidden AI Tool Demo | Humanitarians AI Fellows
A Humanitarians AI Fellow demos Gemini Live's free screen-sharing feature, using it to walk through Google Sheets conditional formatting and live VS Code debugging.
Most AI chat tools still expect you to describe your screen in words. Gemini Live, inside Google AI Studio, skips that step entirely: you share your actual screen, and the model responds to what it can see in real time. This demo, built by a Humanitarians AI student fellow, tests that capability on two very different everyday tasks, spreadsheet formatting and live code debugging, to see how well it actually holds up.
Getting started with Gemini Live
Accessing the feature requires nothing more than a Google account. From Google AI Studio, selecting "stream real time" and choosing a video source lets you start screen sharing directly with Gemini. The video notes upfront that this is still an experimental feature, which sets expectations for the rest of the demo: this is a preview of where the tool is headed, not a fully polished product.
Automating conditional formatting in Google Sheets
The first task shared is a spreadsheet with an "end date" column, and the goal is to automatically highlight any dates that have already passed relative to a specific reference date, without manually checking every row. Gemini confirms it can see the shared screen, correctly identifies the Google Sheet, and restates the request back before offering help, a small but important step that confirms it actually understood what was being asked rather than just pattern-matching keywords.
Gemini walks through the fix step by step: select the entire end date column, open Format, choose Conditional Formatting, and under "format cells if," a rule can be built either with a custom formula comparing dates or with the simpler built-in "date is before" option. When asked directly whether to use the custom formula or the simpler built-in rule, Gemini explains the trade-off, custom formulas offer more flexibility and control, but the simpler option was chosen here, and Gemini adjusts its guidance accordingly rather than insisting on one approach.
Adding a color gradient
Building on the same spreadsheet, a second request asks for a green gradient across a "duration days" column, with darker green for higher numbers and lighter green for lower ones. Gemini again restates the request to confirm understanding before walking through the steps: select the column, add another conditional formatting rule, and choose a green color scale from the built-in gradient options. The gradient applies automatically once the color scale is selected, with no manual color-by-color adjustment needed.
Switching to code debugging
The second half of the demo moves to a completely different kind of task: debugging a Python script for what's described as an AI research project success predictor. After deliberately introducing an error, removing an import and messing with a variable, the same live screen-share session is used to ask Gemini for help, testing whether its usefulness carries over from spreadsheet work into actual code.
Gemini correctly identifies the visible error, missing import for LinearRegression from scikit-learn, and walks through adding the correct import line: from sklearn.linear_model import LinearRegression. When a typing mistake produces a still-underlined error, Gemini catches that Python is case-sensitive and that the class name needs to match exactly, correctly diagnosing that "linear regression" had been typed instead of "LinearRegression." It also flags a missing set of parentheses needed to actually instantiate the class rather than just reference it.
Catching a second bug on request
Asked to check the rest of the code for additional issues before running it, Gemini finds a second, separate problem: a mismatched variable name, where the code attempts to print a test-set accuracy value under one variable name while it had actually been calculated and stored under a different one earlier in the script. That catch is notable because it required tracing a variable's name across two different lines of code, not just recognizing a familiar error pattern.
A sanity check on test size
One more exchange stands out. When the script sets test_size = 2, Gemini doesn't just accept the value; it explains that test_size is typically a decimal between 0 and 1 representing a proportion of the dataset, or alternatively an integer representing an exact count of data points, and it asks directly whether a test set of only two data points was really intended. That prompts a correction: the actual intent was a 20% test split, so the value gets changed to 0.2. That kind of clarifying question, catching a plausible-looking but likely-wrong value rather than just executing the request literally, is a meaningfully different behavior than a purely reactive coding assistant.
Where the tool still falls short
The fellow presenting the demo is candid about its limits: despite these successful catches, Gemini Live is not described as very good at debugging code overall in its current, experimental state, though the expectation is that it will improve. The framing throughout is optimistic but grounded, treating this as an early look at a capability worth watching rather than a finished, reliable debugging partner.
Key takeaways
- Gemini Live, accessed through Google AI Studio, lets you share your screen and get real-time AI assistance without installing extra software.
- It successfully walked through Google Sheets conditional formatting, including a date-based highlight rule and a color gradient scale.
- In a live coding session, it correctly diagnosed a missing scikit-learn import, a case-sensitive naming error, and a separate mismatched variable name bug.
- It asked a clarifying question about an unusual
test_size = 2value rather than silently accepting it, catching a likely mistake before code was rerun. - The feature is still experimental and, per the fellow's own assessment, not yet strong at debugging overall, though expected to improve.
- Access requires only a Google account and is free to use through Google AI Studio.
Who this is for
This demo is built for anyone curious about hands-on, real-time AI assistance for everyday spreadsheet and coding tasks, and it's a good starting point for beginners who want a guided, conversational alternative to searching documentation. It comes from the Humanitarians AI Fellows Program, which pairs a learn-by-doing approach with a learn-by-teaching one, letting student fellows deepen their own understanding of AI tools by demonstrating them for others.
Full transcript(auto-generated, with timestamps)
[0:02]Hi guys, this is Google AI Studio and now you can share your screen and get help from Gemini. All you have to do is just search for Google AI studio and once you open it, you have to go to stream real time. You can just click select video source and start sharing your screen. So I have a spreadsheet over here and I'm going to just ask for a few highlighting help and also have a code here a VS code though it's not like it's still in an experimental version we are just going to try to get as much as help from it right so let's give it a
[0:40]Shot going to stop sharing my Hi, Gemini. Can you hear me? >> Yes, I can. What would you like to talk about? >> Well, can you see my screen now? >> Yes, I can see your screen. It looks like you have a Google sheet open. Yeah. So you know what I want to do now? Uh can you see this end date column? I want to like highlight the dates which has passed uh the current date just which is March 18th. But I want it to be like automated. I don't want to like manually check one by one. Uh imagine when I have like a lot of data, it's
[1:35]Going to be pretty hard, right? So can you help me on it? I just want it to be the right color as well. Yeah. Okay, so you want to highlight the dates in the end date column that are in the past, specifically compared to March 18th, and you'd like to automate this so you don't have to manually check each one and have them highlighted in red. Is that correct? >> Yep. >> Okay, I can help with that. Do you want to know how to accomplish this using conditional formatting in Google Sheets? >> Yeah. Can you just guide me a little bit? How can I do it?
[2:08]>> Okay, get started. First select the entire end date column which is column C. Then in the top menu click format and select conditional formatting. A pane will open on the right. Under format rules select format cells if and then select custom formula is. In the formula field enter the following equals C one less than date 2025 318. >> Wait I see uh date is before. Can we just use this or should we go with custom? >> Yes, you could use date is before and then enter the date you want to compare it to March 18th, 2025. But using custom formula gives you more flexibility and
[2:50]Control over your formatting. Which one would you prefer? >> I think I'll just go with date is before. And uh now I have to like change the color. >> Okay. Okay. So now that you have selected date is before, click on the formatting style section where it says default. You can now select the fill color you want to use such as red from the color palette and then click done. Right. Thank you. Uh it looks pretty good. And also uh it would be nice if we have like a green gradient over here or any color gradient like a darker color for the highest number and a light color
[3:29]For the lowest number. Okay. So now you'd like to add a green gradient to the duration days column making the cells with the highest numbers darker green and the lowest numbers lighter green. Is that correct? >> Yeah. >> Okay. Let's do that. First, select the entire duration days column, which is column D. Then, in the conditional format rules pane on the right, click the add another rule button at the bottom. Under the format rule section, choose color scale and then the green color scale from the available options. You should see the gradient automatically applied. Is this what you're looking for? >> Yeah. Yeah, this is good.
[4:13]>> Great. Is there anything else I can help you with on this spreadsheet? >> No, no. I I think it's good for now. Thank you. >> You're welcome. Let me know if you need any more help. >> So guys, like you see, I just used Germany for this and it it works pretty good. and for our next task what I'm going to remove is I'm going to like remove few inputs and write like add issue right here and maybe increase the test size something like this and let's ask uh Gemini's help let's check if it's able to help us this time let's choose a different voice.
[5:21]>> Hi, Gemini. Can you see my screen? >> Yes, I can see your screen. >> You can see my screen? Yes, I can see your screen. It looks like you are working on a Python script for an AI research project success predictor. Is there anything you would like me to do with this information? >> Yeah, can you see I have an error over here? Uh, model equals linear regression. Can you help me on that? >> Yes, I see the error. It seems like the regression function call is not recognized. You likely need to import the linear regression class from the scikitlearn library. Did you want me to
[5:56]Suggest how to import it? Yeah, can you please help me like how do I put it? >> To import linear regression from scikitlearn, you should add the following line at the beginning of your code along with your other import statements from scarn.linear_model import linear regression. Would you like any further assistance? >> Wait, give me a second. Let me just type in what you said. So, am I writing it right? Yes, you're on the right track. After from scr.linear_model import a, you need to specify what you want to import. In this case, it is linear regression. Therefore, the full line should read from scur.linear_model import linear regression. Would you like
[6:42]Me to confirm it once you've written the line? >> Yeah. Uh, that's it. And I'll still get get an underline over here. Do I need to like I see. >> I don't know. The underline under linear regression indicates that you need to create an instance of the linear regression class before using it. You can do that by adding parenthesis to the end like this linear regression. So the complete line should look like model equals linear regression. After making that change, the underline should be removed and the code should run. Would you like me to explain? >> Can can you see that it's still underlined here?
[7:20]>> Yes, I can see that linear regression is still underlined. Did you add the parenthesis at the end like this linear regression? If it is still underlined. Okay, it seems you have typed linear regression instead of linear regression. Python is case sensitive. So you need to make sure that the class name matches exactly. >> Oh, got it. Thank you for that. And can you check for any other errors in this code before I run it? If there is anything, can you just point that one out? Okay, I see a potential issue on line 32. You are trying to print the test set accuracy, but it looks like you are
[7:58]Referring to it as accuracy instead of accuracy, which is what you calculated in line 27. Would you like me to point out any other issues? >> Okay. So, you're saying uh other than that, everything is good here. >> Yes, that is correct. The rest of the code appears to be good. Do you have any other questions about it? I'm just going to run it and check if it's able to run. >> Sounds good. Let me know if you have any other questions or need further assistance. >> So, uh the model did run uh fine in the Python script, but have a doubt here. So, in this line, can you see the test
[8:35]Size is assigned as two. Is that right? >> Okay, I see the doubt you have in line 19. Test underscore size equals 2 would split the data into train and test data. The test set would have two data points while the rest of the data points go in the training data set. Is that what you meant? >> Yeah. So, is it supposed to be two? Are you sure about that? >> Usually, test size is a decimal between 0 and one representing the proportion of the data set to include in the test split or an integer that represents the exact number of elements to include in
[9:10]The test set. A value of two is valid, but it will result in a small test set when compared to the rest of your data. Are you sure you want a test size of only two data points? >> No, I actually want 20%. >> Okay. To set the test size to 20%. You should change test underscore size equals 2 to test underscore size equals 0.2. Do you have any other questions before you run the script again? >> No. Uh, this is good. Thank you. >> You're welcome. Let me know if you need anything else. >> Right. So guys, here is it. Here is it is the
[9:47]Results and I think >> Okay, >> one second. Gemini, thank you for helping me out. >> You're very welcome. I'm glad I could assist you. Is there anything else I can help with? >> No, thank you. See, it's really good. And the highlight is it's still in an experimental version and imagine once it's fully developed what it can do. uh it can really help you in you know getting you adopted to a new software. Uh it's not going to be like you're using a software for first time when you have an assistance. It's like someone sitting next to you and guiding you how to use it. And uh let me know in the
[10:29]Comment sections what software you think we should be uh getting help from uh by Gemini and maybe we could try it out. And I think it's not like very good at debugging the codes for now, but I'm pretty sure it's going to get better. Thank you guys for watching this video and I would really suggest everyone to give it a shot because you don't have to pay for this. You just need a Google account and you can just log in and try this one out. Thank you.
More videos
2:08Bridging the Pixel Gap in Browser Automation.
2:23How One Narrow Safety Rule Can Make an AI Less Safe Everywhere Else.
2:04Why splitting a chunk from its document makes it retrieve for the wrong question
4:20Three You Can Take Back. One You Can't.
2:21Why a 50-turn agent pays for the same screenshot 35 times unless it caches the pixels
1:53