> ## Content Index > Fetch the complete content index at: https://saadiq.xyz/newsletter/llms.txt > Use this file to discover other available public pages before exploring further. # Let AI Do the Clicking - URL: https://saadiq.xyz/newsletter/let-ai-do-the-clicking/ - Published: 2026-11-11T18:13:49.000Z - Updated: 2026-11-11T18:13:49.000Z - Description: Somewhere a team is paid to make transferring, cancelling, or deleting a little harder than it should be. An AI model can now click through all of it for you. Here's when that's worth doing, how to turn it on in Claude and ChatGPT tonight, and what to do when one of them refuses. - Author: Saadiq Rodgers-King - Tags: Tips, AI Workflows, AI If you've ever had a web app side project, or even just the dream of one, you've bought some domains. I have a pile of them, for apps I built, apps I meant to build, and a few family names. Lots of places to buy domains aren't great places to manage them. I've wanted mine in one place for years, and I kept not doing it. Working through the process is tedious. A domain transfer is something you babysit, potentially for days. The authorization code can take a day to show up. The old registrar can sit on the release for five more. Every wait is a chance to forget where you were across multiple domains. What happens if I miss an email or lose track of the process? On Monday it occurred to me that Claude could do the routine for me. I had seven domains to move to Cloudflare. Five sat at Squarespace, which inherited them when it bought Google Domains, and two at GoDaddy. Claude worked in my own Chrome window, using my logged-in accounts. It copied the DNS records to Cloudflare first so nothing would go dark, then changed nameservers on three domains, unlocked six, and requested their transfer codes. The seventh is staying put for now. Claude checked Cloudflare's list of supported extensions and found that it doesn't take .nyc domains. Claude even polled my email for the authorization code emails. The GoDaddy domain that could move finished overnight. The other five are waiting for Squarespace to release them. The Claude session transcript shows how annoying the process is. An offer to "Protect My Domain." An offer to form an LLC. A bundle of four more extensions for $17.23\. An estimate that my domain is worth $311\. Claude shows a short label for each step as it works, and the labels scrolling by read "Selects 'No longer needed' reason," then "Declines retention offer and continues cancellation," then "Declines loyalty offer." At one point it explained a delay by noting that the site "adds retention screens, so this one takes a few extra clicks." I am infinitely tickled by this. There are whole teams whose purpose is to make it just a little less likely that you do the thing you came to do. Somebody designed each of those screens. I only know they were there because Claude mentioned declining them. It calmly said no and moved on to the unlock button. ## Two capabilities, one idea **Browser use** means the model operates websites in a browser, including the ones you're signed into. **Computer use** goes further. The model takes screenshots of your desktop and moves the mouse and keyboard in any app. I described it to a friend as the AI app sock puppeting your whole computer. Both exist for the same reason. Most software is built for people to click, not for programs to call. When there is no API, these functions allow the model to interact like a person would. ## When it's worth it Usually it isn't. Clicking is the slowest way a model can work, because every step is a screenshot, a decision, and a click. When an API or a built-in connector exists, use that. Half of my domain move ran that way. Claude created the Cloudflare zones and copied the DNS records over the API in seconds. I reach for sock puppet mode in two situations. The first is a chore spread across sites I'm logged into, where the UI is built to slow me down. Registrar transfers are the model case. I could have done the Cloudflare side by hand too. Signing in to both registrars and letting the model click would have been enough. The second is a laborious UI with no bulk option. A friend was handing a work Granola account back to an employer. It held personal notes and 1:1s nobody expected a boss to read. Granola's trash keeps deleted notes for 30 days. To delete one permanently you open it, click a trash icon, and confirm. There were 457 of them. ## What happened when I tried I started by asking Claude to find an API. There's no public one. It found a private API inside the app bundle, and when it tried to decrypt the login token, the model's safeguards stopped it. So I switched to computer use in the Claude desktop app. It wouldn't delete anything. Flatly. I tried telling it that I had backups of the files so it wasn't so much deleting as it was helping me reorganize. It pretty much told me, "That's great. Still, no." Then I opened the Codex desktop app from OpenAI and gave it the same job. It opened the trash, deleted a note, confirmed, and repeated. After a few rounds it scripted the loop. Twenty-nine minutes later it reported all 457 processed, reloaded Granola to check that the deletions stuck, and noted it hadn't touched anything outside the trash. I'm not telling you this to rank the two. In my experience Claude can be more conservative. Lots of times that's a feature. Claude in Chrome always asks before a permanent deletion, and it won't make a purchase or create an account no matter what you approve. It can also feel a little patronizing when it blocks something you consider safe. I don't mind it tapping out if I ask it how to make chemical weapons (hypothetically, of course), but if I want to hack my TV, I don't need Claude shaking its finger at me. These products change constantly, and each one has its own quirks about what it will and won't do. The intelligence underneath is commoditizing fast, though. Claude could have cleared that trash as easily as Codex did. It just wouldn't. When one model refuses a job you know is legitimate, don't spend twenty minutes negotiating with it. Open another one and give it the same prompt. ## Set it up tonight Menu names below are from each company's documentation this week. Both Claude features need a paid plan. ### Claude in a browser 1. Install the Claude extension from the Chrome Web Store, sign in, and pin it from the puzzle-piece menu. 2. In the Claude desktop app, click your initials in the lower left and open Settings, then "Customize," then "Connectors." Find Claude in Chrome, click "Configure," and switch it on. 3. In a new conversation, open the "Connectors" drop-down and enable Claude in Chrome. It's off by default for each conversation. 4. Leave approvals on for your first run. Claude will ask before acting on each new site, and "Allow this time only" is the cautious answer. Running it from the desktop app means the same conversation can click through a site, search your Gmail, and use any other connector you've set up. You can also skip the desktop app and click the Claude icon in Chrome to work from a side panel. ### Claude on your whole computer 1. In the Claude desktop app on Mac or Windows, open Settings. Look under "This computer," then "System." If that isn't there, look under "Capabilities." 2. Turn on "Enable computer use." 3. On a Mac, grant Accessibility and Screen Recording when prompted. It asks before using each app, and the approval lasts for that session. Browsers are view-only in this mode, so use the extension for websites. Available on Pro, Max, and Ultra plans, not Team or Enterprise. ### ChatGPT on your whole computer 1. In the ChatGPT desktop app, switch to Work or Codex and open Settings, then "Computer use." 2. Under "Control," find "Any App" and install it or switch it on. 3. On a Mac, grant Screen Recording and Accessibility. 4. In a chat, type `@Computer` or `@` and the app's name. It asks before using each app. It won't operate terminal apps or approve system security prompts. OpenAI limits this feature to certain regions. ### ChatGPT in a browser 1. In the same "Computer use" settings, expand "More browsers," pick yours, and install the extension. 2. Start a Work or Codex chat and choose the browser from the `@` menu, for example `@Chrome`. It asks per site, with "Allow once" as the cautious choice. Atlas, OpenAI's standalone browser, was retired in August, so this extension is the current route. ## The first run Pick a chore where every step can be undone or sits behind an approval. Then give the model a bounded job. ```markdown I want to move my domains from [old registrar] to [new registrar]. Both are open and signed in. First, list every domain you find and tell me which ones can transfer. Do not change anything yet. After I approve the list, proceed with the transfers. Stop and ask me before any step that exposes personal information, costs money, or can't be undone. Skip every offer and upgrade. When you're done, give me a table of what you changed on each domain. ``` You don't need to spell out the steps, or even know them. The model will work out the process from the pages in front of it. The model you pick matters more than the wording. I ran the domain move in Claude with Opus. In Codex, I'd choose Sol or Astra. Astra is what cleared the Granola trash. A few operating rules I'd keep. - Do the logins yourself. Claude stopped at Squarespace's passkey prompt and at GoDaddy's password page and waited for me. I'd want that from any of them. - Do the payment yourself. I pasted the codes into Cloudflare and checked out by hand. - Ask for the plan before the clicks and the report after. - Stay in the room the first time. ## It can still take a wrong turn A model driving on your behalf will sometimes go somewhere you didn't intend. A year or two ago the stories were comically horrific, with agents deleting databases nobody asked them to touch. The permission prompts are much better now, and so are the models. It's still buyer beware. So start with the chore you've been putting off because the website is miserable. Sign in, keep approvals on, and watch it work once before you trust it with the next one.