beginner 45 minutes

Claude scheduled tasks: get an AI research digest by email

Use Claude scheduled tasks or Gemini Workspace Studio to email you a research digest, with a date window, trusted sources and a dated archive. No coding needed.

A hand-drawn illustration of a yellow robot at a desk of newspapers handing a summary sheet to a green dog in a postal cap with a DIGEST satchel, under a calendar with Thursday and Sunday circled, beside an open filing drawer labelled ARCHIVE.
You'll be able to
  • write a research prompt with a date window, a list of sources you trust and a time limit;
  • set up a scheduled run that saves a dated copy of each report and delivers the summary to you, by email where your assistant can send one;
  • check a digest’s dates and sources before you act on it;
  • find another job you repeat each week that your assistant could run on a schedule in the same way.
Who it’s for
anyone who follows a topic for work and wants a regular summary without checking the news every day. No coding.
Time
about 45 minutes for the first setup and a test run (estimated), then a few minutes after each of the first two scheduled runs to check them.

It’s Monday morning and you have ten minutes before your first meeting. You open the news to catch up on new AI model releases, and half of what you read is the same story you saw on Friday. Checking the news every day takes time, and if you skip a week you can miss a launch your team is already discussing. So how do you get a digest of the news that matters delivered to your inbox on a schedule, without checking every day and and without repeating stories you’ve already read?

For my own research digest on new AI model releases, I’d schedule it twice a week: Thursday and Sunday, ready in the morning. That way I can flip through it during 10 or 20 minutes of downtime when I’m not at my desk. If I have things scheduling every single day, then I’m just recycling the news. If X launches Grokbot or Facebook launches Muse, I want time to process that: to see the reactions, what it looks like in production and what the security implications are, before I get deep into the details.

The Gemini Workspace Studio and Claude scheduled-task routes below have both been tested. The table after them lists what the other assistants offer, from their help pages. The method itself is the same whichever one you use.

Part 1 of 4 Learn

The method: a research librarian and a post person

The job is: go, check the news, analyse it, put it through the lens of what I need, and then give me the summary back. That’s two jobs. I think of it as a research librarian and a post person working together. The librarian finds and reads the sources. The post person turns the result into a summary and delivers it.

You can run them as one flow or two. Splitting them gives you reusable parts: the research chain can be used for other things too. But there’s no reason it can’t be one flow, as long as you’re confident troubleshooting one flow. Don’t engineer it past that point. In Gemini the two jobs are separate steps, and in Claude one prompt can do both.

A six-stage flow diagram, left to right: schedule; topic, date window and time limit; research from trusted sources; a dated archive; a summary; and an email. The first stages are bracketed as the research librarian and the last two as the post person.

Whichever you choose, the method has six parts:

  1. A schedule. Pick the days and the time you’ll actually read it.
  2. A topic, a date window and a time limit.
  3. The research. Ideally from a list of sources you already trust.
  4. A dated archive copy of the full report.
  5. A summary written only from the archived report, with web search off.
  6. Delivery of the full summary in the email body.

The schedule (part 1) is set when you build the flow in your assistant.

Step 1: Write the question and its limits

A research run with no limits can get out of control, running up token costs (a token is a small chunk of text, and assistants measure your usage in them). Give it a date window and a time budget.

A date window. The run should cover only the days since the last one. This matters because it’s a repeatable, reusable flow: we don’t want it to run to the very beginning of time every single time. You want a hard start and a hard stop. In Workspace Studio, the schedule gives you the run’s start time as a chip (a variable you drop into a later step), so the prompt can cover the right dates without you editing it.

How often the run happens sets how wide that window is.

Uses the first line of the guide’s Step 1 prompt. The dates come from today’s date on your device.

Set the date window

Pick the days your digest runs. You’ll see the hard start and stop dates for each run, and which days get missed or covered twice.

Run days
Each run covers

Starts on the day of the last run and stops the day before this one.

The window the guide uses for your first test run.

Covers the one day before each run.

How each day of the week is covered

  1. Mon
  2. Tue
  3. Wed
  4. Thu
  5. Fri
  6. Sat
  7. Sun

Your runs in the next seven days

RunStartStopPrompt

These dates are right for that one run. A scheduled run needs dates that move with each run: in Workspace Studio, build them from the schedule’s Start time chip (a variable from the schedule step that you drop into the prompt).

This page sends nothing anywhere. Your choices stay in your browser.

How to read this

A scheduled digest should cover only the days since the last run, with a hard start and a hard stop, so it doesn’t go back to the beginning of time every run.

This page starts each window on the day of the last run and stops it the day before this run, so each day is in exactly one digest.

A day with no run covering it is a gap: a launch that day won’t be in any digest. A day covered by two runs is an overlap: the same story can turn up twice.

Just because I can be updated every day or every week doesn’t mean that I should be.

A time budget. When you’re starting out, tell it not to spend more than two or three minutes on the task, or five, or ten. This shows you how long the research gathering actually takes, and what blocks it when things don’t work the first time. Check one news source and see how long that takes, then a second one. Testing in small chunks means you can troubleshoot each one.

Paste this into any assistant and change the topic. It uses a five-minute limit:

Research new AI model releases announced between [START DATE] and [END DATE].
Use only these sources: [YOUR SOURCE LIST].
Spend no more than 5 minutes.
For each release: what was announced, by whom, on what date, and a link to the original announcement.
Don't invent anything. If there is no evidence for a claim, leave it out.
If nothing was released in the window, say so.

Step 2: Keep a list of sources you trust

You can keep your sources in a Google Sheet. That way the research and analysis come from sources you’ve already checked and trust.

Just because the information is shouted about loudest doesn’t mean that it’s accurate.

Any spreadsheet or document works. Add a column for when each source was last checked. None of the assistants’ help pages say they will stick to a source list, so put the list in the prompt and check the citations in what comes back.

Tip

In Gemini, attach the Sheet to the step instead of pasting its URL. A bare Sheet URL failed in testing, and attaching it worked. Reattach it after you edit the list.

Step 3: Archive each run with the date in the name

Put the run date at the start of the filename, so each run keeps its own copy of the full report.

The archive is cheap: it’s just text. It’s useful when you need to go back and re-reference something. Company one comes out with a new feature, but company two did that three months ago, so you can go back and check the trend. Over time you can compare old reports with what actually happened.

Step 4: Put the summary in the email

The first thing that broke for me was getting a link to the research back instead of the analysis. Coming back with links that you then have to check kind of defeats the purpose of the automation. You want it all in the email, in one place, without clicking through to anything.

The fix was to add steps. I had to lengthen the automation so it would shorten the process for me to read the results. A separate summary step reads the full report and writes the analysis, and only that text goes into the email.

Two panels. On the left, an envelope holds only a link while a stack of unread papers waits beside it. On the right, the envelope holds a full summary sheet. An arrow between them reads "Add a read step and a summary step".

Ask the summary step for:

  • the date window it covered;
  • the main takeaways;
  • what changed, item by item, with the source link;
  • why each item matters to you;
  • one question worth following up.

Turn web search off in this step, so it uses only the report.


Doing it in Gemini Workspace Studio (tested)

This is the route I built and tested, in a Google Workspace account. Workspace Studio comes with Google Workspace business, enterprise and education plans, and personal accounts can use it through Google Workspace Experiments; check Google’s Workspace Studio help page for your account before you start. The flow has six steps:

  1. On a schedule: set the days and time. I’d pick Thursday and Sunday mornings.
  2. Deep Research: the prompt from Step 1, public web on, Workspace off, with the schedule’s Start time chip in the prompt for the date window.
  3. Ask Gemini: reads the complete report through the Link to Deep Research report variable from step 2. Deep Research only outputs a link, so this step turns it into text.
  4. Create a doc: from the text step 3 produced, with the Start time chip at the start of the filename. This is your dated archive.
  5. Ask Gemini: writes the summary described in Step 4 above, working only on the text from step 3, with web search and Workspace off.
  6. Notify me by email: put the Content created by Gemini chip from step 5 in the Message field.

Six numbered cards in a row, left to right: on a schedule (Thursday and Sunday); Deep Research, which gives a link; read the report, so the link becomes text; create a doc, a dated copy; write the summary, with web search off; and email me, with the summary in the body.

Use Test run before you turn the schedule on. It takes real actions: it sends a real email to the account you’re signed in with, and Notify me by email only ever goes to the person running the flow. Test each step on its own, then the whole chain.

Warning

Writing “send mail” in a prompt doesn’t send anything. The email only goes out through the Notify me by email step with the Content created by Gemini chip in it.

Doing it in Claude (tested)

Claude’s cloud scheduled tasks (Anthropic called them routines when I tested this, and its docs now say “formerly routines”) can do the whole job in one prompt, because the Gmail and Google Drive connectors let it send the email and create the Doc itself. In a test run on 18 October 2026, it researched the window, saved a dated Google Doc, and sent an email with the full summary in the body, in about three minutes and with no permission prompts.

  1. In Claude Code, type /schedule and describe the task: a name, the days and time, and the prompt. You can also use claude.ai/code/routines. You don’t need a repository.
  2. Paste the prompt from Step 1 into the task, then add three instructions: create a Google Doc with the run date at the start of the title and the full report in it; write the summary from that report only; and send one email with the Gmail connector, to your own address only, with the summary as the body text and the Doc link at the end.
  3. Check the connector list in the task’s settings. Every connector you have is included by default. Keep Gmail and Google Drive and remove the rest. In my test, asking /schedule to remove the others didn’t work, and all of them stayed attached.
  4. Use Run now once, and check the email and the Doc before you leave the schedule on.

Scheduled tasks are on Pro, Max, Ultra, Team and Enterprise plans and are in research preview, so the names and screens may change (Anthropic’s docs, checked 5 November 2026). Claude also has Desktop scheduled tasks, which run on your own computer instead of in the cloud.

Warning

A scheduled task uses connector tools without asking, including sending email. Tell it in the prompt who the email goes to, and make that only you.

Doing it in other assistants

Every assistant below can run a prompt on a schedule. The last column shows what each one sends by email. I re-checked the Claude, Microsoft 365 Copilot and Gemini app rows against their help pages on 5 November 2026. The ChatGPT row was last checked on 26 October 2026.

AssistantScheduling featurePlansWhat arrives by email
ChatGPTScheduled tasksIncludes Free: 3 active tasks, at most once a day, in a morning, afternoon or night window. Paid plans get more tasks and exact timesA push or email notification. With the Gmail app connected, a task can send email, but sending may pause until you approve it. OpenAI’s help page doesn’t say whether the notification email includes the full result
ClaudeScheduled tasks, with the Gmail and Google Drive connectorsPro, Max, Ultra, Team and EnterpriseThe full summary, in the body (tested)
Microsoft 365 CopilotScheduled promptsNeeds a Microsoft Copilot licence. Up to 10 promptsAn optional email notification “when the response is ready”. The response itself is in Copilot’s Chats list. Microsoft’s help page doesn’t say whether the email includes the response
Gemini appSchedulesPersonal accounts, rolling out; qualifying Workspace editions. Up to 10 activeAn app notification. Google’s help page doesn’t mention email

Note

If your assistant only sends a notification, you can still keep Steps 1 to 3: run the saved prompt on its schedule, then copy the result into a dated document yourself. On a free account with no scheduler, set a calendar reminder for your two days and paste the saved prompt in by hand.

Check it before you trust it

A successful run only shows that the email arrived. You still need to check the content. In my tests, some outputs gave imprecise dates. Before you act on a digest, check:

  • Dates: each item falls inside the window.
  • Sources: each link opens and says what the summary claims. I open every link before I trust an item.
  • Secondhand reports: if an item links to a news story, find the company’s own announcement and check it says the same thing.
  • “So what” claims: the analysis can draw conclusions the sources don’t support. Check it against the sources before you repeat it.

Use the same method for other repeated jobs

The same method works for other jobs you repeat every week. Write down what you did in the first thirty minutes of each of your last three workdays, and look for anything that shows up more than once with the same steps each time. For example, “Check my email and flag anything that needs a reply” uses the same steps each morning on different emails.

Once you have a candidate, start with something simple with clear inputs and outputs. For a first task, try “Summarise my unread email”. A job that files attachments and updates your CRM has too many steps to test at once.

Run it as an ordinary prompt on a few different days before you schedule it, and adjust it each time:

  • What it gets wrong: add a constraint, such as “only the Work calendar”.
  • What it misses: add a step.
  • What you don’t need: tell it to leave it out, or to use one line instead of a paragraph.

When the output is useful every time, schedule it the same way as the digest. Two to start from:

Morning email summary: check my unread email from the last 24 hours. For each message, give me who it's from, one line on what they want, and whether it needs a reply (yes, no or FYI). Put replies needed first and newsletters last. Skip promotional email. Don't reply to anything.
Friday week review: look at the documents I edited in the last 7 days. Tell me what I spent most time on, what got finished, what's still in progress, and anything untouched for over a week. Under 200 words.

These need your assistant to be connected to your email or files. Check which connections your scheduled run can use before you rely on it.

Part 2 of 4 Practise

Try it on your own work

  1. Pick a topic you follow for work and the two days you’d actually read about it.Result one topic, two days and a time, written down.
  2. Write your source list: three to five sources you already trust, in a sheet or document.Result a short list with a “last checked” column.
  3. Fill in the Step 1 prompt with your topic, a date window covering the last seven days (for this first test run; the schedule itself runs twice a week), your sources and a five-minute limit, and run it once by hand.Result a list of items, each with a date inside your window and a link to the original. If nothing happened, it says so.
  4. Build the research step in your assistant’s scheduler, with your prompt and the days and time from step 1.Result a scheduled task or flow that runs your prompt, and a test run that returns the same kind of list you got by hand.
  5. Add the dated archive and the summary email, if your assistant can send one.Result a test run that leaves a document with the date at the start of its name, and the summary in an email. If your assistant only sends a notification, the summary is in the assistant and you copy it into the dated document yourself.
  6. Check the test run with the list under “Check it before you trust it”.Result you can say which items you’d act on and which you’d drop, and why.
  7. Write down one other repeated job from your last three workdays.Result one job where the steps stay the same, ready to test as a prompt.

Part 3 of 4 Check

Check yourself

Your Thursday digest arrives with an item dated nine days ago. Your date window was the last seven days, as for a first test run. What do you do?

Leave the item out until you’ve checked it, then look at the prompt. The run went outside its window, so check that the prompt gives a hard start date and a hard end date for the window. In Workspace Studio, build the dates from the schedule’s Start time chip so they move with each run.

Your Workspace Studio flow sends you an email with only a link to the research report. What do you check?

Check two things. First, the read step and the summary step: Deep Research only outputs a link, so you need an Ask Gemini step that reads the full report into text, then a summary step working only on that text. Second, the email step: its Message field should hold the summary step’s Content created by Gemini chip. If it holds the report link, swap it.

A digest says a company "has changed the market" with a new release, and links to a news story. Can you repeat that in a meeting?

Not yet. Find the company’s own announcement and check it says what the news story says. A line about why an item matters is the summary’s own reading, so check it against the sources before you repeat it.

You want your Claude scheduled task to email the digest. What do you check before you turn the schedule on?

That the prompt sends the email to your own address only, that only the Gmail and Google Drive connectors are attached, and that one Run now produced the email and the Doc you expected. A scheduled task uses its connectors without asking.

Your assistant only sends a notification when a scheduled run finishes, with no email. How do you still get a digest you can read and keep?

Run the saved prompt on its schedule, read the result in the assistant, and paste it into a document with the run date at the start of its name. You keep Steps 1 to 3 and do the delivery by hand.

A colleague wants to schedule "do my morning routine". What would you suggest instead?

Pick one repeated job with the same steps each time, such as summarising unread email, write it as a specific prompt, and run it by hand on a few days until the output is useful. Then schedule it.

Part 4 of 4 Keep

Keep this

Start with one schedule

  • Run the digest twice a week, with a date window, a list of sources you trust and a dated copy of each report.
  • Have it draft, not send. If a scheduled job produces something for other people, have it save a draft for you to review. Only the digest to yourself goes out by itself.
  • Each run counts against your usage limits. Start with one or two schedules and check your usage before you add more.
  • One schedule first. See whether the output is useful before you build more.

In the Kit Library, members can get the research prompt and the two scheduled prompts as one file, ready to paste into Claude or Gemini.