01The Afternoon I Lost to a Hand Icon
It started, as these things do, with the assumption that I had broken something.
I run Claude for Chrome, the browser extension that lets Claude actually do things on a page instead of just talking about them. Click buttons, read content, fill fields. It is genuinely useful right up until the moment it decides to ask your permission for every keystroke, and then it is the most exhausting tool you own.
That is where I found myself on Facebook. Every action, a prompt. Read the page? Prompt. Click a thing? Prompt. A little hand icon at the bottom of the side panel that read "Ask before acting," and a tooltip that, when I hovered it, told me with total serenity: "This site requires permission for each action."
My immediate, confident, completely wrong conclusion: I did this. I fat-fingered some setting during a session and now I have to undo it.
So began the hunt.
02The Reinstall Reflex
When software misbehaves, the lizard brain reaches for the oldest ritual we have: turn it off and on again. Uninstall, reinstall, surely that wipes whatever bad state I created.
I uninstalled the extension. I reinstalled it. I opened Facebook with the smug confidence of a man about to be proven right.
Every action. Still a prompt.
This is the moment the afternoon tips from "quick fix" into "something is wrong with the universe." Because a reinstall is supposed to be the great reset, the scorched-earth option. If that did nothing, then whatever was happening was not living in some local cache I could nuke. It was coming from somewhere I could not reach. I just did not yet understand where, or why.
03The Tour of Every Settings Panel That Could Not Help Me
Here is the embarrassing middle of this story, the part where I went looking for a lever that did not exist and was so sure it must.
I dug into the extension's own settings: icon, three dots, Settings, Permissions. There is a real page there. It has a section literally called "Your approved sites," which is exactly the thing you would want when your problem is a site behaving wrong. Mine said, helpfully:
No sites have been approved yet.
Empty. Not Facebook-with-a-bad-setting. Empty. You cannot revoke what was never granted, and you cannot, from this page, add a site either. The whole panel is a list of past decisions, not a place to make new ones.
I checked Chrome's own extension site-access controls. I checked my account settings on the web. I kept hunting for the toggle that turns "ask every time" into "stop asking," the way every other permission system on Earth lets you do.
There is no such toggle. And the longer I looked for one, the more I was just generating plausible-sounding places it might live and then being disappointed they were empty. That is its own kind of frustrating, the search that keeps feeling like it is one click from done.
04The Screenshot That Ended the Argument
The thing that finally told me the truth was the permission prompt itself. I triggered an action on Facebook and actually read the box instead of swatting it away.
"New permissions required." Claude wants to read page content on: www.facebook.com. Two buttons: Allow this action. Decline.
And then, in small grey text at the bottom, the sentence that retroactively explained my entire afternoon:
Site-level permissions are disabled for this site.
There it was. Not "you misconfigured this." Not "approve this and we'll remember." On a normal site, that prompt offers a third option, "Always allow actions on this site," the one that adds it to your approved list and shuts off the nagging. On Facebook, that option is gone. Deliberately removed. The fine print even spells out the philosophy: Claude will not purchase items, create accounts, or bypass captchas without your input.
Facebook is on a list. A list of sensitive sites, social media, banking, email, where the extension refuses to let you hand over the keys, no matter how much you want to. You never set it. You cannot unset it. There is no it to unset.
05Except Gmail Already Solved This Exact Problem
Here is the thing that turns my frustration from "annoying" into "indefensible," and it is the whole reason I am writing this instead of just sighing and clicking Allow.
Claude does not behave this way in Gmail. Not even close.
In Gmail, Claude reads my inbox, understands a thread, writes a complete, well-formed reply, and drops it straight into my Drafts, all of it, without stopping to ask permission for each step. It does the entire job. The one thing it will not do is hit Send. That single irreversible act it hands back to me, every time, and I am glad it does.
That is the correct model. Sit with how good it is. Everything reversible, Claude just does. The one act that cannot be taken back, it gates. A draft is harmless. It sits there, I can edit it, delete it, ignore it forever. Sending is the point of no return, so sending is where the human belongs. One gate, placed exactly on the thing that matters.
Now hold that up against Facebook. Reading the page, reversible, harmless, the digital equivalent of looking, gets a prompt. Clicking around, prompt. The graceful, graduated model that Gmail nails is simply absent. Facebook treats "look at this page" with the same suspicion it would treat "broadcast a post to everyone you have ever met."
06The Gate Is in the Wrong Place
This is the actual bug, and it is not a labeling problem like I first wanted to believe. It is that the gate is bolted to the wrong layer.
Facebook absolutely has irreversible actions: publishing a post, sending a DM, firing off a friend request, leaving a comment. Fine. Gate those. Gate them exactly the way Gmail gates Send: let Claude compose the post, draft the message, set the whole thing up in the box, and then stop, hand me the irreversible click, and wait. I would love that. That is the contract I already trust in Gmail.
Instead, Facebook gates reading and navigating, the harmless, reversible majority of what an agent does, with the same friction as the dangerous minority. And that is not just annoying, it is quietly less safe. When a tool makes me click "Allow" forty times an hour for actions that obviously do not matter, "Allow" stops being a decision and becomes a reflex. I am rubber-stamping. So on action forty-one, the one that does post publicly, my finger is already moving toward Allow before my brain catches up. Prompt fatigue doesn't add safety. It trains it out of you.
Anthropic already built the better pattern. It ships in the same extension, one product over. They know the right answer is "do everything reversible, gate the irreversible commit," because that is precisely what they did with email. Facebook just doesn't get the good model, and nobody will tell you why.
07What I Actually Want
I am not asking to disable a safety feature. I want to be exact about that, because the moment you complain about an AI guardrail somebody assumes you want the guardrails gone. I don't.
I want the Gmail contract applied to Facebook. Let Claude read the page. Let it navigate. Let it draft a post or a reply right into the compose box. And then let it stop at Publish, at Send, at the one button that cannot be un-pressed, and ask me there, and only there.
That is not less safety. It is better-placed safety: one sharp, meaningful gate on the irreversible act, instead of a blunt prompt-on-everything that exhausts me until I stop reading the prompts at all. Anthropic already proved they can build it. I just want them to admit Facebook deserves the same respect my inbox gets.
08What To Actually Do (So You Don't Lose Your Afternoon)
Until that day, here is the short version I wish someone had handed me:
- You did not break it. Stop trying to fix it. The per-action prompt on Facebook (and other sensitive sites) is enforced, not configured.
- Reinstalling will not help. It is not local state. Save yourself the ritual.
- There is no toggle. Not in the extension, not in Chrome, not in your account. The "Always allow on this site" button is intentionally absent on protected sites.
- The only move is Allow this action, every time. That is the design, for now.
- Check the fine print on the prompt. "Site-level permissions are disabled for this site" is the tell. When you see it, you are not misconfigured. You are on the wrong side of a design decision.
I started this afternoon thinking I had a settings problem. I do not. I have a design problem, and it is not the one I expected. The fix was never going to be a better-worded tooltip. A clearer sign on a locked door is still a locked door. The fix is the model Anthropic already built and shipped in the next product over: let the agent do all the reversible work, and put the one gate where it belongs, on the irreversible act. Gmail lets Claude write the whole email and stops it at Send. Facebook should let Claude read, navigate, and draft the post, and stop it at Publish. That is not me asking for less safety. It is me asking them to move the seatbelt from my wrists to the one place a crash actually happens. They know how. They've done it. Facebook is just still waiting for its turn.