Your Weekly Report Only Needs You Once. After That, It Writes Itself.

You have written the same report forty-two times. It has never once needed your judgment, and it has never once been fun.

The Report That Eats Your Tuesday

Every small business has one. A weekly sales summary. A monthly client recap. A stock check before the Friday order goes in. A cash position pulled from three different accounts before the Monday standup. None of it is hard. All of it is the same four moves, done from memory, over and over: pull the numbers, compare them to last time, decide what actually matters, write it down in plain words, send it.

Eliot Prince knows the shape of that problem better than most. Before he started teaching AI workflows on YouTube, he ran an SEO agency, and he says client reporting cost him a full week every month: logging into Google Search Console, exporting the data, comparing this month to last year to strip out seasonality, writing the commentary, then drafting the email to send it. Multiply that by a dozen clients and you have lost a quarter of your working month to a task that never once required a new idea, only patience.

That exact shape, pull it, compare it, explain it, send it, is what a feature Anthropic shipped in July is built to eat.

Claude Learned to Watch Instead of Listen

On July 21, 2026, Anthropic added Record a Skill to Claude Cowork, its desktop workspace mode. It sits under the plus menu, and it is available on the Pro, Max, and Team plans, not the free tier. The mechanic is different from anything Claude shipped before it. Instead of writing a prompt describing what you want done, you click record, do the task yourself on screen, and talk through it as you go. Claude watches your clicks, your typing, and your narration, then turns the whole demonstration into a reusable skill it can run again without you.

Anthropic's own example for this feature was almost word for word a reporting workflow. As PCMag summarized it: if you manually run a regular report that collates data from multiple sources, you can show Claude where to find each piece of information and how you want it presented, and the system hands you back a repeatable skill you can turn on to run the report. That is not our framing. That is the use case Anthropic itself reached for first.

The wider pattern matters too. OpenAI shipped a similar screen-recording feature called Record and Replay for Codex about a month earlier, in June, though it launched Mac-only. Two of the largest AI labs converging on the same idea inside a single month is a signal worth taking seriously: teaching an agent by demonstration, not by describing, is becoming the default way small teams will hand off repetitive work.

Build the Report Agent: Six Steps

Here is the build, based on how Eliot Prince walked Claude through his own SEO report, generalized so it works for any recurring report your business already produces.

1. Pick the report you already write from memory. Not your hardest task, your most repetitive one. A weekly sales recap, a stock count, a job-cost summary, a client status update. If you could do it half asleep, it is a good candidate.
2. Open Claude Desktop, switch on Cowork, and choose Record a Skill from the plus menu. You need the desktop app for this, since Claude may need to drive your browser for parts of the task.
3. Do the report once, on camera, narrating it like you are training a new hire. Say what you are clicking and why. Say what counts as a good week versus a bad one. Say how you phrase the comparison, and what tone you use when the numbers are ugly. This is the only time you will ever have to explain it.
4. Answer Claude's follow-up questions. After the recording, Claude asks what it was unsure about. In Prince's case: should it pull the source data through browser automation, what format should the final report take, and what should happen with the outgoing email. He chose browser automation, a branded PDF, and draft-only, meaning Claude prepares the email but never sends it without him.
5. Test-run it and fix what reads wrong. Run the new skill once. Check the numbers against what you would have written by hand. If the tone sounds generic instead of like you, or your branding is missing, edit the skill directly rather than starting over.
6. Put it on a schedule, or bind it to a command. Once it is producing work you would sign your name to, set it to run weekly or monthly on its own, or leave it as a one-word command you run when you remember.

Same Loop, Different Reports

The mechanism does not care what industry you are in or what system holds your numbers.

A freelance marketer in Austin can record her own monthly client-performance recap, the one she currently rebuilds from three dashboards and a spreadsheet, and hand the whole recurring job to a skill that reads like her, because it learned from watching her do it.

A plumbing firm in Manchester can teach Claude its month-end job-cost report, the one that currently means pulling numbers from a job-tracking app, a supplier invoice folder, and a bank statement, then reconciling them by hand before the owner sees a true profit number.

A small retailer in Nairobi can build the Friday-night ritual most shopkeepers already do without calling it a report: checking stock against what sold that week and reconciling it with M-Pesa takings, and turn that half hour of squinting at two screens into something that runs itself.

A logistics dispatcher in Lagos can automate the weekly delivery summary that used to mean collecting a dozen driver text messages by hand and typing them into one sheet before the client call.

Different tools, different currencies, same shape: pull it, compare it, explain it, send it. Show it once, and the shape is what gets saved, not the specific screen you happened to be looking at.

Three Ways This Can Go Wrong

None of this is free of risk, and treating it as magic is how it burns you.

First, a skill built by watching a screen is only as stable as that screen. One automation vendor pushed back publicly on this exact feature, arguing that UI-driven agents are more brittle than ones built on APIs, because a redesigned dashboard or an unexpected pop-up can break the workflow a human never would have noticed breaking. They are not wrong. Check your new skill's output against reality for the first few runs before you trust it blind.

Second, keep a human in the loop until the skill has earned your confidence. Prince chose to have Claude draft the client email rather than send it automatically, specifically so a person reviews the numbers before a client or investor sees them. Do the same with anything that leaves the building.

Third, do not narrate your passwords. Log in before you hit record, or use whatever connector your accounts support, so the only thing Claude learns is the workflow, not your credentials.