Schedule a Recurring Report With AI That Reads Your Files

To schedule a recurring report with AI that actually reads your files, first work out where the files live. If they are in Google Drive, a connected app or files you've uploaded to the AI tool, a cloud schedule works: Claude Cowork's scheduled tasks, ChatGPT tasks on the web, or Claude Code's cloud routines. If they sit in a folder on your Mac, the schedule has to run on that Mac, and it only fires while the Mac is awake with the app open. The fix is not a cleverer prompt. It is a computer that stays on.

This guide covers what each scheduler really does as of September 2026, a decision table, how to write the job so it produces a file you can use, and how to keep it running when your laptop is shut.

What does "scheduled AI task" mean in Claude, ChatGPT and local agents?#

The same phrase covers two different things, and most guides blur them.

A cloud schedule runs on the vendor's servers. It fires at nine whether your laptop is open, asleep or in a bag. The price is that it sees only what the vendor can reach: connectors, files you uploaded, a fresh clone of a repository.

A local schedule runs on your computer. It can open the folder where you keep last month's reconciliation, the export your bank emails you, and the Excel model nobody has moved to the cloud. The price is that your computer has to be on and the app running when the clock strikes.

Here is how the main tools split, taken from their own documentation:

• Claude Cowork. Anthropic's help center says scheduled tasks "run remotely, so they run on their cadence even when your computer is asleep or the Claude Desktop app is closed." It also says they "work with your connectors and the files saved to your Claude account. They can't be tied to a folder on your computer." And: "If a scheduled task requires local files or apps, it will only run locally." Cowork scheduling is on every paid Claude plan.
• Claude Code Desktop. It has both kinds, with the tradeoff spelled out in a table. A local task has access to local files but "only fires while the app is open and your computer is awake." A cloud routine runs with the computer off but works on a fresh clone, not your files, and can run at most once an hour.
• ChatGPT. OpenAI's scheduled-tasks documentation says tasks in the desktop app can work with local projects, and to "keep the computer on and the app running when a scheduled task needs local files." Web tasks "can use uploaded context and connected tools, but they can't work directly in a folder on your computer." Scheduled tasks are on Plus, Pro, Business and Enterprise.

Neither company hides this. The tradeoff is written into their own docs.

Why can't a cloud AI schedule see my files, and why does a local one stop?#

The cloud side. A remote job has no route to your disk, and giving it one would mean syncing your folders to someone else's servers. Plenty of recurring reports live in exactly the places that never get synced: a bank CSV in Downloads, a QuickBooks export, a client's rent roll in .xlsx, a folder of PDFs a scanner dropped on a network drive. A cloud task pointed at that work can fail, or write a report from whatever it could reach, which is worse.

The local side. A local scheduler is an app checking the clock. Anthropic's Claude Code Desktop docs are blunt about it: "If your computer sleeps through a scheduled time, the run is skipped." Turning on Keep computer awake stops idle sleep, but "closing the laptop lid still puts it to sleep." When the Mac wakes, Claude Code Desktop starts one catch-up run for the most recently missed time within the last seven days and drops the older ones. So your Monday 8am report might run at 7pm on Tuesday, when you open the lid.

For a report somebody reads first thing Monday, a late run is a missed run.

Which scheduler should I use? A decision table#

If your row is one of the first three, stop reading and use the cloud scheduler you already pay for. Everything below is for people whose work is on a disk.

How to schedule a recurring report with AI: write the prompt, test it, then schedule it#

The mistake is scheduling a prompt you have never watched run. Do it in this order.

1. Write the prompt as a standing instruction, not a chat message#

A scheduled prompt runs with nobody there to answer questions. Write it the way you'd brief a contractor who starts on Monday:

• Inputs, by path. "Read every CSV in ~/Clients/Acme/bank/ dated in the last seven days, and invoices.xlsx in the same folder." Don't write "the bank stuff."
• The date window, stated as a rule. "The week is Monday 00:00 to Sunday 23:59 before today." A late run then still covers the right week. Anthropic's docs give the same advice: build timing guardrails into the prompt itself.
• The output, as a file with a name. "Write reports/2026-W38-reconciliation.csv and a one-page summary.md beside it." More on this below.
• What to do when something is missing. "If there is no bank file for this week, write a summary that says so and stop." Otherwise the agent may make one up.
• What not to do. "Don't change any existing file. Don't send anything."

2. Run it by hand, twice#

Run the prompt as a normal session and read what comes back. Then run it again against a different week. The first run shows whether it understood you. The second shows whether it is repeatable, and repeatable is the whole point.

While you watch, note any permission prompts. In Claude Code Desktop, a task running in a mode that needs approvals stalls until someone approves. Anthropic suggests pressing Run now after creating the task and choosing "always allow" for each tool it needs.

3. Then schedule it#

Pick a time with slack. If the report is read at 9am, schedule it for 6am. Say "weekly on Monday" rather than "every 7 days," so the day doesn't drift when a run slips.

How do I make the AI's output a file instead of a chat message?#

A chat reply is a report nobody can find next week. A file can be opened, compared with last week's, attached, and fed into the next run.

Three habits make the difference:

1. Name files by period. 2026-W38-pipeline.csv, not report.csv. The agent can then read last week's file and say what changed, which is usually the most useful line in the report.
2. Tables go in CSV or a sheet, prose goes in markdown or a doc. Ask for both: the rows you sort and filter, and five sentences on what matters in them.
3. Never let it overwrite the model. If the numbers feed an Excel workbook with formulas, have the agent write an input sheet or a separate CSV and link to it yourself. We cover that in how to get AI to fill in an existing Excel model without breaking formulas.

For a weekly deck, the same rule applies. Ask for a new file each week, not edits to the master.

How do I keep a Mac running for scheduled AI tasks? caffeinate, pmset or a spare Mac mini#

If the schedule has to be local, there are three levels of effort.

Level 1: Keep this Mac awake. For a desktop Mac, open System Settings, turn off automatic sleep while plugged in, and leave the app running. For a MacBook on power, caffeinate -s in Terminal blocks system sleep while that command runs, and caffeinate -i blocks idle sleep. Neither survives closing the lid, unless the MacBook is running closed with power and an external display attached.

Level 2: Wake it on a timer. sudo pmset repeat wakeorpoweron MTWRF 05:55:00 tells macOS to wake or power on the Mac at 5:55 on weekdays, a few minutes before a 6am job. It is one line, it survives reboots, and pmset -g sched shows what is set. It still needs the app to be running, or to be opened, when the Mac wakes.

Level 3: A machine whose only job is staying on. A Mac mini, or last generation's laptop with the lid open on a shelf. Set it to restart automatically after a power failure, never sleep, and sign in on its own. This is the only setup where "every Monday at 6" actually means every Monday. We walk through the full setup in run AI agents overnight on a spare Mac.

A spare machine also solves a problem that the awake settings don't: you can shut your own laptop without thinking about the schedule.

How do I check a scheduled run happened, and what do I do when one is missed?#

Silence is the dangerous failure. A schedule that stops quietly is worse than none, because you stop checking the thing you delegated.

• Look at run history, not just the output folder. Claude Code Desktop keeps a history per task and shows skipped runs, with the reason on hover: the computer was asleep, the previous run was still going, or other tasks were already running.
• Make the report prove its date. Put "Covers 8–14 September, generated [timestamp]" on the first line. A stale file then announces itself.
• Diff against last week. "What changed since 2026-W37" catches a run that read the wrong folder, because the diff comes out absurd.
• When a run is missed, run it by hand for that period. Because the prompt defines the week relative to today, a manual run on Tuesday covers the right days. Don't let two runs overlap on the same output file.

Five weekly reports worth automating#

These all share one shape: the inputs change every week, the method doesn't, and a person reads the result.

1. Bank reconciliation. Last week's bank lines matched against issued invoices, with the unmatched ones listed with amounts. Bookkeepers can see the method in how to reconcile a bank statement against invoices with AI and on the bookkeeping page.
2. Sales pipeline. Deals that moved, deals that went quiet for fourteen days, and next steps that are past due, read from a CRM export or a tracker sheet.
3. Listings watch. New and changed listings on the streets or in the price band you care about, as a dated table. Agents and brokers can start from the real estate page.
4. Competitor changes. Pricing pages, changelogs and job posts from a fixed list of companies, with only the changes called out. See automating competitive analysis.
5. Weekly KPI pack. The same six numbers from the same sources, a short paragraph on each that moved, and last week's file beside it. Agencies running this for several clients can see the marketing agency page.

Where Universe fits#

Universe is a Mac app that runs agents on the model account you already pay for: Claude through Claude Code, ChatGPT through Codex, Gemini, Grok, or open models on your own machine. It doesn't resell model usage, and there are no credits.

For this job, the relevant parts are:

• Scheduled work runs on a Mac. You set up a recurring job ("do this every weekday at eight") and it runs at that time whether or not you're looking. On a Mac, Universe can register a wake so the scheduled job runs.
• The output is a real file. Sheets, documents and decks land in the session's folder, and the panel shows each one as what it is. Up to 24 earlier versions of each file are kept on the Mac, so if a run overwrites last week's report, the old one can be restored.
• A second Mac keeps it going. You can link a spare laptop or the machine under your desk, and the agents keep working after you close the lid.
• Schedules, the browser and other Macs are on the free plan. Plans only change amounts; see pricing.

The limits, stated plainly:

• It is Mac only. Apple silicon or Intel, macOS 13 or later. There is no Windows, web or phone app.
• Schedules don't run in the cloud. A schedule can miss if its Mac is off. The spare Mac has to actually be on.
• A turn moved to another Mac has less to work with. It has the model's own tools, but not Universe's own tools such as browser control. Plain files still sync back. Browser profiles also can't move between Macs. If the report needs a signed-in browser, run Universe on the spare Mac itself, with a Chrome profile copied on that machine.
• It won't send the report for you. Sending, paying and posting wait for a person, so a scheduled run leaves a file and a draft, not an email already sent.

If your inputs are all in Drive and connected apps, a cloud schedule in Claude Cowork or ChatGPT is simpler, and you don't need a second machine. Universe is for the case in the middle of the table: the files are on a disk, and the report still has to show up on Monday. Download it here. For reports that read sites behind your own login, see letting an AI agent work on sites behind your login.