There is a system with no export button, a portal that will not talk to anything else, an old piece of software you are stuck with. So you sit there and click. Copy a value, paste it, tab across, repeat, fifty times. It is the kind of task that has resisted every automation tool because there was nothing to connect to.

Claude can now do that clicking for you.

Not through an integration or a connector. Through your actual screen. It looks at what is in front of you, moves the mouse, clicks the buttons, types in the fields, the same way you would. If it can be done by a person at a keyboard, it can now be handed off.

This week: the specific work tasks worth handing to an AI that drives your computer, and the model that just launched with a memory big enough to swallow an entire book in one go.

The Old Way:

Every business has at least one tool that traps you. An old CRM the owner will not replace. A government or supplier portal with no export. An internal system built a decade ago. A piece of industry software that has never heard of an API.

Automation tools could never touch these, because there was nothing to plug into. Zapier needs a connection. A script needs an interface. These systems offer neither. So the work stayed manual, and it stayed yours. You copy a number from one screen, paste it into another. You update fifty records one at a time. You pull the same report every Monday by clicking the same eleven buttons. Nobody chose this. It was just the work that could not be automated.

The Replacement:

Claude can now use your computer the way you do.

When it does not have a direct connector for a task, it takes over your screen instead: it looks at what is there, then points, clicks, scrolls, and types to get the job done, asking your permission before it acts (source: Anthropic). It opens files, drives your browser, and works through the software already on your machine, with no setup (source: Anthropic release notes).

This is what unlocks the trapped work. The old CRM with no export, Claude can read it and type into it, because it uses it like a person. The portal with no integration, same thing. The task that was manual purely because there was nothing to connect to is now a task you can hand off.

This is a research preview, not a finished product. It launched earlier this year and has been improving since. The genuinely new part this month is that you can now trigger it from your phone: message the task from your mobile and Claude does it on the Mac or PC sitting on your desk.

What this replaces: the click-heavy grind in the tools that no other automation can reach.

What it costs: a Claude Pro or Max plan you may already have. It is included, no extra tool.

The thing that makes this matter is not that it is clever. It is that it reaches the work everything else could not.

Every automation tool until now needed a door: an API, an integration, a connector. That left a whole category of work untouched, the stuff locked inside systems with no door. For most businesses, that untouchable category is exactly where the worst grind lives, because the clunky old tools are the ones nobody built integrations for.

Computer use ignores the door. It walks up to the window and uses the software the way a human does, by looking and clicking. That is slower than a direct integration, and Anthropic says so plainly: working through your screen takes longer than a clean connection, and complex tasks sometimes need a second try (source: Anthropic).

The Pro section includes:

🔒 Pro members get all 6 tasks with copy-paste briefs, the briefing method, the safety rules, the honest guide to where it breaks, and the one-page cheat sheet. Put the trapped work down this week.

xAI shipped a tool that builds a working voice agent in under two minutes, no code required (source: AIApps).

A voice agent answers your phone, holds a real conversation, and handles the call: booking an appointment, answering common questions, taking a message, qualifying a lead. The old way to get one meant a developer or an expensive platform and weeks of setup. This builds one in the time it takes to make a coffee.

The pricing is genuinely small: $0.05 per minute of audio plus $0.01 per minute for telephony (source: AIApps). A ten-minute call costs about 60 cents. For any business that misses calls (and the data says most small businesses miss a lot of them), an agent that picks up every time pays for itself fast.

🔒 Implementation checklist is for pro members. Get a breakdown plan to implement AI into your workflow.

💸 A cheap open model just landed inside GitHub Copilot. On July 1, Copilot added Kimi K2.7 Code from China's Moonshot AI, the first open-weight model in its picker, pitched as a lower-cost option at roughly $0.95 per million input tokens (source: The Stack).

🔒 Meta's new image tool shipped with a privacy default worth checking. Muse Image launched with an opt-out setting that lets others pull your public Instagram photos into their AI generations, and it drew a backlash within hours (source: The Bridge Chronicle). Why this matters for you: if you or your business post public photos on Instagram, take two minutes to check your AI settings so your images are not used in ways you did not choose.

📱 A capable AI model now runs entirely on your phone, offline. PrismML released Bonsai 27B, a model that fits in under 4GB and runs on an iPhone with no internet, handling text, images, reasoning, and code, free under an open license (source: Future Tools). On-device AI means private, offline, no-subscription options are becoming real, which matters for anyone handling sensitive data or working where connection is patchy.

🏢 Google launched a way to run whole fleets of AI agents, not just one. At Cloud Next '26, Google unveiled Gemini Enterprise, a platform for building and governing many agents that connect to your data and run multi-step workflows under IT control (source: Build Fast with AI). It is Google's direct answer to ChatGPT Work and Claude Cowork, which means the big three are now all racing on business agents. Competition like this pushes capability up and price down for you.

On July 17, Google released Gemini 3.5 Pro with a 2-million-token context window, the largest memory in any production AI model, double what most rivals offer.

The number is abstract. What it means is not. Two million tokens is roughly 1.5 million words. You can now hand an AI an entire book, a full codebase, or a year of documents in a single request and have it reason across all of it at once (source: Zoombangla).

To see why that matters, look at what people have been doing without it. For two years the real limit on AI was not intelligence, it was memory. The window was too small to hold much, so you fed it pieces. You split documents. You pasted one file at a time. You summarised things before asking about them, and you watched the model forget the beginning of a long input by the time it reached the end. Every clever workaround people built existed to get around one wall: the AI could not hold enough at once.

That wall just moved a long way back. When a model holds your whole contract library, it can find the one clause that appears across all of them, something impossible when it could only see one contract at a time. When it holds a year of support tickets, it can spot the pattern that only shows up across hundreds of them. When it holds an entire codebase, it sees how the whole thing fits together instead of one file.

There is an honest catch, and it matters. A big window is not the same as reliable reasoning across all of it. Independent testing through 2025 found that every model degrades as the input grows, and often well before the advertised ceiling (source: TechTimes). So 2 million tokens does not mean flawless recall of 2 million tokens. It means you can now attempt things that were flatly impossible before, while still checking the important answers.

The shift for you is in what you now ask. Most people are still using AI as if the window were tiny, feeding it one page at a time. The advantage this month goes to the people who realise they can stop chopping things up and start handing over the whole pile.

🔒 The full breakdown is for Pro members: the specific "load the whole pile and ask" workflows this unlocks, three copy-paste prompts to run them, how to prepare a big pile so the model can actually navigate it, and how to check an answer drawn from documents you have not read yourself.

Workflow Library Updates:

NEW Resources:

Each resource lives permanently in your Pro account. Use them whenever you need them.

Every business has work that got left behind by automation, the click-heavy grind trapped in tools with no way in. That work was manual because there was no other option. This month, there is one. It is rough, it is a preview, and it still reaches further than anything that came before.

If you do one thing before next Thursday, pick the most tedious trapped task you have and hand it to Claude's screen. Watch it once. Then let it grind so you do not have to.

👇 What’s behind the paywall:

Pro members deploy AI in their work an average of 5-8x more often than free readers (based on reply data from past issues). The difference is having the exact setup, not the concept.