\n
Most people open Gemini Spark<\/u>, type something into it the way they would type into a chatbot, and come away unimpressed. That is the wrong mental model. Spark is not a better chat window. It is closer to hiring an assistant who works overnight, and the difference between a useful result and a useless one is almost entirely in how you brief it.<\/p>\n
This guide walks through what Spark is, how to set up your first task, the three building blocks everything else is made of, the prompt pattern Google\u2019s own examples all follow, and the four things Spark still cannot do.<\/p>\n<\/p>\nTable of contents<\/p>\n
<\/default:svg><\/button><\/p>\n\n1. First, do you have it?<\/li>\n 2. When to use Spark instead of ordinary Gemini<\/li>\n 3. Getting started<\/li>\n 4. How to write a Spark prompt that works<\/li>\n 5. Level one: your first task<\/li>\n 6. Level two: make it proactive<\/li>\n 7. Level three: teach it a skill<\/li>\n 8. Level four: chain a skill to a schedule<\/li>\n 9. Letting Spark use your browser<\/li>\n 10. Ready-made prompts to copy<\/li>\n 11. What Spark cannot do yet<\/li>\n 12. Staying in control<\/li>\n 13. Spark versus the alternatives<\/li>\n 14. What is coming next<\/li>\n<\/ol>\n<\/section>\nFirst, do you have it?<\/h2>\n\nHow to get started with Gemini Spark.<\/p>\n
Photo: Google<\/p>\n<\/figcaption><\/figure>\n
Spark is already live for Google AI Ultra<\/u><\/b> subscribers across Asia Pacific,<\/span> and is rolling out to Pro<\/u><\/b> subscribers in Singapore over the coming weeks. It works in English and local languages.<\/p>\n It runs on Google\u2019s cloud rather than your device, which is the detail that makes the rest of it possible. Spark keeps working when your laptop is closed and your phone is locked. It integrates with Gmail, Drive, Docs, Sheets and Calendar natively, with no setup required, and it is available on the web and on Mac desktop.<\/p>\n
Under the bonnet, it runs on Gemini 3.6 Flash<\/b>, paired with the same engine behind the Antigravity harness<\/u><\/b>, Google\u2019s coding agent platform. That pairing is why Spark can keep working on a single job for twenty or thirty minutes instead of giving up after one reply. Spark first appeared at Google I\/O in May,<\/u> running on Gemini 3.5 Flash<\/u>.<\/p>\nWhen to use Spark instead of ordinary Gemini<\/h2>\n\nKnowing when to use Spark is important.<\/p>\n
Photo: Google<\/p>\n<\/figcaption><\/figure>\n
A quick rule of thumb, because this trips up most new users.<\/p>\n
Use ordinary Gemini chat<\/b> when you want an answer now, when you are thinking out loud, or when you want to go back and forth on an idea. It is a conversation.<\/p>\n
Use Spark<\/b> when the job has any of these characteristics:<\/p>\n\n
\nIt will take longer than a single reply, such as producing a long document or a full presentation<\/li>\n It needs to happen repeatedly, on a schedule<\/li>\n It needs to happen in response to something, such as an email arriving<\/li>\n It involves acting on your own Gmail, Drive, Docs, Sheets or Calendar<\/li>\n It needs a browser to go and do something on a website<\/li>\n<\/ul>\n<\/div>\nAdam Coimbra, Director of Product Management on the Gemini team, calls this category \u201cdigital laundry\u201d, the repetitive work that pulls attention away from what actually matters.<\/p>\n
\u201cJust some task that kind of is a distraction of your attention from the more important things,\u201d he said, \u201cthat are where you actually want to be able to focus.\u201d<\/p>\n
Getting started<\/h2>\n\nGetting Spark to run on a Schedule.<\/p>\n
Photo: Google<\/p>\n<\/figcaption><\/figure>\n
Go to gemini.google.com and look for the switch to Spark. Flipping it changes the interface from a chat box to a task dashboard, with separate sections for schedules and skills.<\/p>\n
That dashboard is home. Every job you give Spark appears there as a task, and you can open any task to watch its reasoning as it works.<\/p>\n
Everything in Spark is built around the following three things. Learn these and the rest follows:<\/p>\n
\n
\nTasks<\/b> are the basic unit. One task is one goal Spark is working on for you. Every time you ask Spark to do something, you are creating a task.<\/li>\nSkills<\/b> are how you teach Spark to behave. Writing style, formatting preferences, the way you want a particular type of job handled. You can write a skill yourself, or you can simply tell Spark to read your existing emails and build one based on how you already write.<\/li>\nSchedules<\/b> are the proactive layer, and they come in two flavours. Time-based schedules run at a set moment, such as every Friday at 8am. Event-based schedules run when something happens, such as an email arriving on a particular topic.<\/li>\n<\/ul>\n<\/div>\n The real power comes from combining them, which we will get to.<\/p>\n
How to write a Spark prompt that works<\/h2>\n Handily, we have some examples from Google to get started properly. If you examine them, they nearly all follow the same four-part shape. To get your prompt to work properly, take note of the following:<\/p>\n
\n
\nTrigger.<\/b> When should this happen? \u201cEvery Friday at 8am.\u201d \u201cWhenever a new flight or hotel booking confirmation email arrives.\u201d \u201cEvery month on the 1st.\u201d<\/li>\nAction. <\/b>What should Spark do? \u201cFind local events.\u201d \u201cAudit my digital invoices in Gmail from the last month.\u201d \u201cScan my inbox for customer inquiries.\u201d<\/li>\nDestination. <\/b>Where should the result go? \u201cAdd them to my rolling Weekend Plans Google Doc.\u201d \u201cAdd the details to my My Bali Trip Itinerary Google Sheet.\u201d \u201cSave it to drafts for my review.\u201d<\/li>\nCondition. <\/b>Any rules or limits? \u201cIf there are 1 to 3 events that you highly recommend, go ahead and add them to my calendar.\u201d \u201cIf only outdoor courts are available, alert me only if the forecast is clear.\u201d<\/li>\n<\/ul>\n<\/div>\n Miss the trigger, and you get a one-off instead of an automation. Miss the destination and the work vanishes into a chat reply you will never find again. The condition is optional, but it is what stops Spark from burying you in noise.<\/p>\n
You need less context than you think.<\/b> In the APAC briefing, Coimbra asked Spark to project-manage a studio renovation and deliberately gave it almost nothing. No link to the file. He even used the wrong word.<\/p>\n
\u201cIn fact, I called it a remodel. The actual PDF says renovation. It doesn\u2019t actually say remodel anywhere,\u201d he said. \u201cSo my prompt wasn\u2019t even a good prompt, but it immediately found the file.\u201d<\/p>\n
Spark searched the Drive, found the renovation plan, and rebuilt the entire schedule a month forward, creating a Google Task for each step with details and times, plus calendar events for the major milestones.<\/p>\n
The lesson is to describe the outcome you want rather than laboriously assembling the inputs. Spark will go looking. As Coimbra put it, it \u201cwill automatically understand that you\u2019re a Google user and it\u2019s gonna go find what context it can.\u201d<\/p>\n
Level one: your first task<\/h2>\n Start with something that has an obvious destination and an obvious trigger. A tracker is ideal.<\/p>\n
The demo Google used was an RSVP tracker. The instruction was, in effect, create a spreadsheet to track workshop RSVPs, and log each one as the emails arrive.<\/p>\n
Spark immediately began searching Gmail and Drive for context about the workshop without being told to. It then built the spreadsheet, found previous RSVPs already sitting in the inbox and preloaded them, and set up a scheduled action to fire whenever a new RSVP email arrives, extracting the name, email address, date and status, then appending a row.<\/p>\n
When a test email arrived, Spark woke up on its own, read it, and updated the sheet.<\/p>\n
It also mangled the formatting, which Coimbra logged as a bug on the spot. Worth knowing before you rely on the output looking presentable.<\/p>\n
Level two: make it proactive<\/h2>\n Most AI tools can already run something daily or weekly. \u201cGive me a news update,\u201d or \u201cSummarise the emails I got about this project.\u201d Spark goes further, because you can point it at a subject rather than a time.<\/p>\n
Tell Spark to keep track of a project and it works out that it should wake up hourly, or the moment an email on that topic arrives, and then go and see what has changed. If you use Google Chat, it can read what people are saying across different chat rooms, pull together an update, and maintain a tracker for you in a Google Sheet.<\/p>\n
This is the difference between a reminder and an agent. A reminder tells you to check. An agent checks.<\/p>\n
Level three: teach it a skill<\/h2>\n\nGiving a skill a name allows it to be reusable.<\/p>\n
Photo: Google<\/p>\n<\/figcaption><\/figure>\n
The easiest skill to start with is one Google has already built for you. Open the skills section and look under recommended for Match My Writing Style<\/b>.<\/p>\n
The first time you use it, it reads through your emails and documents and constructs a custom skill around how you actually write. From then on, anything Spark drafts for you sounds more like you.<\/p>\n
You can also build one from scratch by describing it, or by pointing Spark at examples. Google\u2019s own suggested prompt for this is worth copying directly:<\/p>\n
\u201cRead through the last 50 emails that I wrote and turn it into a style guide for how I write emails. Turn that into a skill that gets used every time I ask you to draft emails for me. Call that skill \u2018ghostwriter\u2019.\u201d<\/p>\n
Note the \u2018ghostwriter\u2019 instruction. Naming a skill means you can invoke it deliberately later.<\/p>\n
Level four: chain a skill to a schedule<\/h2>\n\nLet it know when it has to run.<\/p>\n
Photo: Google<\/p>\n<\/figcaption><\/figure>\n
This is where Spark stops being a novelty, and the best example came from Coimbra describing his own setup rather than from any slide.<\/p>\n
He built a skill that reads through all his Google Chats and emails from the past couple of days and produces a list of everything he has promised to do, plus everything other people have promised him. Then he built a schedule that runs that skill automatically and emails him the result.<\/p>\n
\u201cI\u2019m constantly saying yes to things, and yes, I\u2019ll do that, and yes, that sounds good,\u201d he said, \u201cand then I lose track of them.\u201d<\/p>\n
The email he received on the morning of the briefing ran to five pages.<\/p>\n
The pattern generalises. A skill defines how<\/i> to do something well. A schedule defines when<\/i>. Build the skill once, attach a schedule, and the job runs without you ever opening Spark again.<\/p>\nLetting Spark use your browser<\/h2>\n
Spark can drive a browser to complete tasks on websites that have no Google integration at all. In the demo, it was asked to go to Class Bento, a platform for listing local courses, and post a workshop listing for around ten people per session on Wednesday nights.<\/p>\n
Two things are worth knowing before you try this.<\/p>\n
\nIt will ask permission first.<\/b> Spark researched the task by digging through Google Keep and Drive, then stopped and asked before opening a browser at all.<\/li>\nYou get to choose whose browser.<\/b> You can let Spark drive Chrome on your own machine, where you will watch the cursor move by itself, or you can hand the job to its own browser running on a virtual machine. Pick the virtual machine if you want to close your laptop and walk away. Pick your own Chrome if the site needs you to be logged in already.<\/li>\n<\/ul>\n In the demo, Spark got most of the way through the listing, realised it did not have pricing information, and came back to ask rather than inventing an answer. That is the behaviour you want, but it does mean browser tasks are not always set-and-forget.<\/p>\n
Ready-made prompts to copy<\/h2>\n\nSpark allows for deep integrations with other Google apps<\/p>\n
Photo: Google<\/p>\n<\/figcaption><\/figure>\n
Google has published a set of starter prompts<\/u>. These are the ones worth stealing first, and each is a clean illustration of the four-part pattern above:<\/p>\n\n
\nTravel.<\/b> \u201cWhenever a new flight or hotel booking confirmation email arrives, automatically add the details to my \u2018My Bali Trip Itinerary\u2019 Google Sheet.\u201d Then follow up with: \u201cFind interesting local events and activities happening in Bali when I\u2019m there and add them to my calendar on days I\u2019ve got nothing else planned.\u201d<\/li>\nWeekends.<\/b> \u201cEvery Friday at 8am, find local events that you think I\u2019d be interested in and add them to my rolling \u2018Weekend plans\u2019 Google Doc that I can check for inspiration on things to do. Include key info about the events or activities, including links to learn more. If there are 1-3 events that you highly recommend, go ahead and add them to my calendar.\u201d<\/li>\nSport.<\/b> \u201cEvery Tuesday at 8am, log into and check the PickleMe app for free indoor evening courts. If only outdoor courts are available, alert me only if the forecast is clear.\u201d<\/li>\nEvents.<\/b> \u201cFind my wedding RSVPs in Gmail, organise the dietary restrictions into a table, and draft a summary email to our caterer.\u201d<\/li>\nMoney.<\/b> \u201cEvery month on the 1st, audit my digital invoices in Gmail from the last month. Notify me about any subscription price hikes, or trials that are about to end and pre-draft cancellation emails.\u201d<\/li>\nSmall business.<\/b> \u201cScan my inbox for customer inquiries. Automatically draft an email response recommending my services and pricing packages based on their request, and save it to drafts for my review.\u201d<\/li>\nParents.<\/b> Google also suggests pointing Spark at scattered school emails so you never miss another permission slip deadline.<\/li>\n<\/ul>\n<\/div>\nWhat Spark cannot do yet<\/h2>\n Four limitations to know before you build anything important on it:<\/p>\n
\n
\nPhotos and PDFs attached to emails.<\/b> Spark cannot pull them in. Asked about this directly, Coimbra took it personally: \u201cThis is actually a thing I remember deciding to launch, even though we didn\u2019t have this feature. It\u2019s painful for me.\u201d He then committed on the record to fixing it within three weeks of the briefing, which puts it around mid-August 2026. Check whether it has landed.<\/li>\nMultiple Google accounts.<\/b> If your subscription sits on your personal account but your work files live on a separate one, Spark cannot span both. \u201cIt\u2019s like super hard,\u201d Coimbra said. \u201cBut I know it\u2019s a point of friction. I mean, you want this one thing to just manage everything. So I totally get it.\u201d He promised to look at the roadmap.<\/li>\nTelling drafts from final versions.<\/b> If your Drive is full of near-identical versions, Spark may grab the wrong one. Coimbra was candid that this is unsolved. \u201cWe\u2019re at the beginning in the AI era of like how big a problem that\u2019s going to be, because we\u2019re all creating like hundreds and thousands of AI stuff, and then only some of it is real.\u201d For now Spark leans on signals such as whether a document has been shared with other people, on the theory that shared files are likelier to be the real one. The practical workaround is to name your final versions clearly.<\/li>\nUnlimited use.<\/b> Spark comes with your Google AI plan and will not generate API charges, but you do not get unlimited queries. At some point it will tell you that you have run out for now, and access refreshes after a few hours.<\/li>\n<\/ul>\n<\/div>\n