Let ChatGPT Do the Mundane Tasks You Know You Should Be Doing (The 4-Line Prompt)

You know you should be logging every medical receipt. Keeping a record of what the car needed and when. Looking at the numbers behind the emails you send. None of it is hard. It’s just mundane enough that it never quite happens.

That’s the kind of task the ChatGPT desktop app handles best now, and the part of the app that does it is easy to miss. I hadn’t really, either: I’ve used ChatGPT Work for a while and wasn’t sure what it was for.

Receipts changed that. I snap a photo on my iPhone, it syncs to iCloud Drive, and I move it into Downloads on my Windows PC. Then I tell ChatGPT to find it and log it in an Excel spreadsheet in my project folder, following the instructions I saved in the project. The surprise was what started showing up in the log: monthly prescription refills and small medical costs I’d never have bothered to write down. Now they’re tracked, and I can reimburse myself for them tax-free from my HSA.

Speeding up the tasks I already did saved a few minutes. Getting the skipped ones done at all is where the real payoff was.

Below is how to automate tasks with ChatGPT’s desktop app: what ChatGPT Work is, which mundane tasks it handles well, how to pick the right tool with one question, and the four-line prompt that decides whether it finishes the job.

Verified as of September 24, 2026, against OpenAI’s Computer Use and Record & Replay documentation. Features and availability change quickly; check the app before you rely on a detail here.

The quick answer

  • Chat answers. Work does. In the ChatGPT desktop app, Work mode takes a goal, makes its own checklist and carries it out across your files, connected apps, a built-in browser and, with a plugin, other apps on your computer.
  • Start with a task you’ve been skipping. The payoff is biggest on the tasks that never get done because they’re tedious.
  • Where the task lives decides the tool. A folder on your computer: just point Work at it. A website: type @browser. An app installed on your computer: install Computer Use, then type @computer.
  • Four lines do most of the work: where to go, what finished looks like, “if a value is missing or unclear, leave it blank and tell me why,” and “tell me how many you found and how many you processed.”
  • Watch it first. Let it work in the background only after it has done the same task correctly a few times.
  • You make the call on anything that sends, publishes, pays or deletes. Review it first, then give the go-ahead.

What ChatGPT Work is (if you’ve only used the chat)

Regular ChatGPT is a conversation. You ask, it answers, and you do whatever comes next.

Work is the mode in the ChatGPT app that does the next part. You describe the finished thing, and it plans the steps and carries them out: reading the files in a folder you give it, using apps you’ve connected (Gmail, Google Drive, Calendar), opening websites in its own browser, and saving the result as a real file. It’s strongest in the desktop app, because that’s where it can reach your folders and your other apps.

The shift is small: you describe the finished thing, and it does the clicking.

The two kinds of mundane task it handles well

Dylan Davis, who runs an AI consultancy and spent three weeks testing this on client work, sorts the tedious parts of software into two jobs. The split is his, and it’s the clearest way I’ve found to spot a task worth handing off.

Job 1: Moving information. You’re the one carrying data from one place to another: out of a photo and into a spreadsheet, or out of a portal that has no “download all” button, one record at a time. Nothing about it is difficult. It’s just long, which is exactly why it gets skipped.

My receipts are this job. The information sits in a photo, and it needs to end up as a row in a log.

Job 2: Figuring it out. It’s software you rarely touch, where you know what you want but not where it’s hidden: an auto-renewal buried three menus deep, or a report you pull once a year. These tasks get skipped because learning the software feels like more work than the task itself.

For me that’s Payhip, where I sell my Facebook Marketplace flipping guide, and MailerLite, which runs my email list. I log into both just often enough to forget where everything is. I never understood the difference between a MailerLite “campaign” and just sending an email to my list, so I let ChatGPT’s computer use work it out for me. It found its way around, drafted the campaign, and sent it once I’d read the draft and told it to.

(The answer, for the record: a one-off email to your list is a campaign. MailerLite calls it a regular campaign, and the thing to tell apart from it is an automation, which sends on a trigger.)

One question picks the tool: where the task lives

Ask one question before you start. The answer tells you what to type.

Where the task livesWhat to use
A folder on your computer (receipts, downloads, exported files)Point Work at the folder and ask
A website (a portal, a dashboard, an online account)The built-in browser: type @browser
An app installed on your computerThe Computer Use plugin: type @computer
  • A folder is the simplest option. There’s no clicking involved: Work reads the files directly.
  • A website is where most tasks live. Sign in only to the site the task needs.
  • An installed app is the slowest option. On a Mac it works in the background; on Windows it takes over your screen while it runs.

One more rule sits above the table. If ChatGPT already has a proper connection to the app (Gmail, Google Drive and the rest of its connected apps), use that instead. A connector talks to the app directly, which is faster and more reliable than clicking through the screen.

Turning on the built-in browser

The built-in browser comes with the desktop app on both Mac and Windows. Open the side panel from the top-right of the app and choose Browser. It keeps its own sign-ins, separate from Chrome or Safari, so you can sign in there to just the site your task needs.

To hand it a task, type @browser in the message box and pick it from the list. You’ll see it appear as a blue tag, and everything in that message will be done in the browser.

Turning on computer use

Computer Use is a plugin, so it needs installing once. Open Plugins from the left-hand menu, search for “computer,” and install Computer Use. On a Mac, you’ll also be asked to allow screen recording and accessibility access in System Settings. ChatGPT asks your permission app by app the first time it uses each one.

After that, type @computer (or @ plus the app’s name) to hand it a task.

Reach for this one last. Dylan found it slow even on its fastest setting and describes the browser as “night and day” faster by comparison. OpenAI’s own guidance also says to use the built-in browser first for anything on the web.

The 4-line prompt

This is the part to copy. It works for all three places a task can live.

Go to [the folder, website or app].
When you're done, I want [the finished thing, and what shape it should be in].
If a value is missing or unclear, leave it blank and tell me why.
When you're finished, tell me how many you found and how many you processed.

The structure is Dylan Davis’s. Here’s why each line earns its place.

Line 1: Where. The folder, the site or the app. Say what it’s allowed to touch, and add “don’t change any settings” if the task is read-only.

Line 2: What finished looks like. Describe the result you want to end up with. “One row per receipt with date, vendor and amount” gives it a finish line it can check itself against. A list of steps just gives it your guess at the best route.

Line 3: “If a value is missing or unclear, leave it blank and tell me why.” The AI treats its job as giving you an answer. When a receipt is blurry or a field is empty, the easiest answer is a plausible-looking guess. This line makes “I don’t know” an acceptable answer, and a blank cell with a reason is far more useful than a confident wrong number.

Line 4: “Tell me how many you found and how many you processed.” This is the receipt for the whole run. If it found 42 and processed 42, it finished. If it found 42 and processed 39, you know exactly where to look. Without the count, “Done!” tells you nothing.

The rule for reusing it: rewrite lines 1 and 2 every time, and keep lines 3 and 4 on every task, either in the prompt or saved in your project instructions. Lines 1 and 2 describe the task. Lines 3 and 4 are how you check the work.

My receipts, with the four lines

This is the version for my HSA receipt log:

Look in my Downloads folder for receipt photos I've added since [date].
Log each one in my receipt spreadsheet using the project instructions: one row per receipt. Skip any that are already logged.
If a value is missing or unclear, leave it blank and tell me why.
When you're done, tell me how many receipts you found, how many you logged, and which ones you skipped and why.

When I ran it, the count matched: it found every receipt I’d added to Downloads since my last run and logged all of them. My usual, shorter prompt catches them too, because my project instructions already tell ChatGPT what to do with unclear values and to confirm the count.

That points to the best way to use the four lines. They matter most on the first run of a new task, or on a big batch where a skipped item would be easy to miss. Once a routine works, put lines 3 and 4 into your project instructions, and every short “log my new receipts” prompt carries them automatically.

The count is still worth keeping for the day something goes wrong: a photo that didn’t sync, or a receipt saved somewhere else.

The HSA rules let you reimburse yourself tax-free for qualified medical expenses you have after opening the account, and prescribed medicines count. But you have to keep records that show what each withdrawal paid for (IRS Publication 969; IRS Publication 502 covers what counts). A log that quietly skips a refill is a log you can’t fully trust. A log that tells you it skipped one is a log you can fix in a minute.

A website you rarely visit, with the four lines

Here’s how I’d ask my MailerLite question now: read-only, in the built-in browser, plus the numbers I never look at:

@browser Go to MailerLite (I'm already signed in). Don't change any settings or send anything.
Explain in plain English how a campaign differs from an automation, and show me where I'd start a one-off email to my list. Then make me a table of my last five emails with send date, open rate, click rate and unsubscribes.
If a value is missing or unclear, leave it blank and tell me why.
When you're done, tell me how many emails you found and how many you put in the table.

That’s both jobs in one prompt: finding your way around software you rarely use, and moving numbers out of a dashboard.

A website is exactly where the built-in browser beats Computer Use. I used computer use for my MailerLite question the first time; for a site you’re already signed in to, @browser is the faster route.

How much to trust it: the four-rung ladder

Treat it like a new hire. Trust gets earned one task at a time.

  1. Watch every run. Start with tasks that are low-stakes and easy to check, and keep an eye on what it’s doing.
  2. Let it run in the background. Once it has done the same task correctly several times, and the counts keep matching, let it work while you do something else. Spot-check now and then.
  3. Save it as a skill. When the task runs cleanly every time, ask ChatGPT to package it as a skill, a saved set of instructions it can reuse. Test the saved version once, because packaging can break something that worked.
  4. Put it on a schedule. Only then, and only if a schedule makes sense, ask it to run the skill for you: “run this every Monday at 8:00.” Here’s how scheduled tasks work in ChatGPT and Claude.

A schedule is how a task you kept meaning to do becomes one that simply gets done.

If you do the same task fifty times, there’s also a shortcut: do it once while ChatGPT watches, and it writes the skill from your recording. OpenAI calls it Record & Replay. It’s Mac-only for now and not yet available in the EU, UK or Switzerland. I’m testing it against Claude’s version for a separate post.

Where I draw the line on safety

Last year I wrote that I’m waiting on AI browsers. I still am. My concern was an AI acting inside my logged-in accounts, where a hidden instruction on a web page could trick it into doing something I never asked for.

When I wrote about connecting ChatGPT to my bank account, the question I landed on was: can it only read, or can it also act? The bank connection only reads. I’m more comfortable giving ChatGPT read-only access to my accounts through Plaid than letting any AI browse around in my logged-in accounts.

The built-in browser and Computer Use both act. They click and type. That puts them in the higher-care category, so here’s how I use them:

  • Folders first, then connectors. My receipts never touch a website: ChatGPT reads photos in a folder. That’s the lowest-risk version of this whole idea.
  • Sign in to one site at a time. The built-in browser keeps its own sign-ins, so give it only the site the task needs.
  • Never my bank, never my main email. Health portals, like an HSA provider’s site, get the same caution. My HSA task works from photos in a folder for exactly that reason.
  • Send, publish, pay and delete wait for my go-ahead. It can do everything up to that point. When it drafted my MailerLite campaign, it clicked Send only after I’d read the draft and told it to. Who physically clicks matters less than who decides.
  • Stay present for anything sensitive. OpenAI’s own guidance is to keep tasks narrow and stay present for account, security, privacy, payment and credential settings.

Captchas also come back to you. That’s fine: a task that needs you for one click is still a task you mostly didn’t do.

What it still won’t do well

  • Unfamiliar pop-ups. Dylan found that an unexpected pop-up occasionally derails a Computer Use run. Watched runs are how you find out which tasks have one.
  • Speed in desktop apps. Clicking through an installed app is slow. If a task can happen in the browser or in a folder, use that route.
  • Your Windows screen during Computer Use. On Windows it takes over the active desktop, so plan the run for when you’re away from the machine.
  • Judgment calls. It will log a receipt, but it won’t decide whether your pharmacy receipt is HSA-eligible any better than you would. Keep the decisions for yourself; hand it the typing.

The best first task is the one you’ve been putting off: tedious, low-stakes and easy to check. The receipt pile, the car’s service history, the dashboard you never open.

The tasks you keep putting off are the easiest ones to hand over. Start with the one you’ve skipped the longest.

Key takeaways

  • ChatGPT Work is the mode in the ChatGPT app that carries out tasks instead of answering questions.
  • The biggest payoff comes from the mundane tasks you’ve been skipping.
  • Ask where the task lives: a folder (just ask), a website (@browser) or an installed app (@computer).
  • Use the four lines: where, what finished looks like, “leave it blank and tell me why,” and “how many did you find and how many did you process.”
  • Rewrite lines 1 and 2 every time. Keep lines 3 and 4 on every task, in the prompt or in your project instructions.
  • Watch it first, and keep the decision on send, publish, pay and delete for yourself.

Related reading

Leave a Comment