What to do when something has gone wrong, and the questions people ask once they have started. None of it loses your work.
Four files. The Bluestocking Guide is how it works. The Bluestocking Setup Instructions is what you do. Claude setup instructions is the file you upload - it’s for Claude, not for you. This is the fourth: what’s actually broken, and the questions people ask once they’ve started.
Two different things live here. When something’s wrong is for a specific symptom - something that should work and isn’t. Everything after it is the questions people ask once they’re running, in the order they tend to ask them - not steps, not problems, just how the pieces fit.
Cause: A write can report success without landing - a permission that was never granted on a new page, or Claude saying “done” before actually reading the page back.
Do this:
Don’t fix it by hand in Notion. Claude won’t know you did, so its account of what’s on that page stays wrong, and the next thing it writes will be built on the wrong picture. If you already have, that’s fine and there’s an entry for it further down.
You’ll know it worked when: the page shows the content, in Claude’s own account of it, not a summary you have to take on faith.
Cause: Either the connection was never approved for that page - usually from picking pages individually instead of choosing everything during setup - or the connection has dropped since.
Do this: on a computer or browser, not the phone app - connectors can’t be changed from a phone.
You’ll know it worked when: you type can you see my Notion? and it says yes.
Cause: A usage limit or a closed tab ended the session before a stage’s write was confirmed.
Do this:
Don’t start setup over from the beginning. Claude reads what’s already in Notion and builds only what’s missing, so restarting risks duplicate pages, which is a second, worse problem.
You’ll know it worked when: Claude says what it found already built, and starts only on what’s missing, without re-asking anything you already answered.
Cause: Claude finds pages by name, not by folder location. Two pages sharing a name is a coin flip every time.
Do this:
These names need to be unique in your workspace: Home, Log, Roles, Rooms, Projects, Corrections, Check log, Claude Auditor. Notion lets you use them elsewhere and won’t warn you.
You’ll know it worked when: you ask which page Claude’s reading and it names the one you meant.
Cause: Almost always a rename. The lists on a project page aren’t copies - they’re windows that find things by name. Rename the project and every window pointing at the old name comes back empty. Nothing is lost and nothing shows an error. It just looks blank.
The same thing happens if you rename a room, or if a project is called Life Admin in one place and Life admin in another. The capital letter counts.
Do this:
You’ll know it worked when: the page shows your things again. If it doesn’t, nothing was deleted - say so in the same chat and it’ll keep looking.
Cause: Nothing’s broken. But Claude’s picture of that page is now out of date, because it only knows what it wrote. Anything it says about that page until it re-reads it is built on the old version.
Do this:
Editing by hand is allowed and sometimes it’s the fastest thing. Just say so afterwards.
You’ll know it worked when: Claude describes the page the way it actually reads now.
Cause: There is no end of a conversation for Claude to write at. A chat ends when you stop typing, and Claude never gets another turn - so anything meant to happen “at the end” never happens at all. Where I left off is written as you go instead, and replaced whenever what you’d come back to changes. If it’s empty, that write hadn’t happened yet when you closed the chat.
Do this:
Going forward, say your stopping phrase when you have to stop suddenly - the one you chose at the start of setup, listed under Words to Remember on Home. It forces the write there and then rather than leaving it to a moment that never arrives.
You’ll know it worked when: you open a new chat in that room, ask where did I leave off, and it answers from the page rather than asking you.
Cause: Every chat starts by going and reading Home, because nothing is
remembered between conversations. That read is the cost of the memory switch
being off, and it isn’t a fault. It gets worse as Home grows, and worse again
if Home links to something long that keeps getting pulled in with it.
Do this:
more than Home and the current project’s page, something is being read that
doesn’t need to be.
along with an offer to move the old parts of it somewhere else. Nothing is
deleted.
there’s nothing to wait for when you want to throw something at it. Read the
next entry before you rely on that.
You’ll know it worked when: a new chat answers without a long pause, and
what it says it read is Home plus the project page.
Cause: They can’t. Not while they’re open, and not afterwards.
That’s the cost of turning the memory switch off, and it’s the trade you made
on purpose: nothing about you is quietly accumulating inside Claude, so every
chat starts knowing nothing and has to go and read Notion. Two chats open side
by side are two strangers reading the same filing cabinet.
What that actually means:
written to Notion. If the first chat hasn’t written it yet, the second has
never heard of it - and it won’t say so, because it doesn’t know there was
anything to miss.
line, and the second one overwrites the first. Your log is safe, because
those are separate rows. That one line isn’t.
If you changed Home this morning, that chat is still working from the old
version.
Do this: one chat at a time for anything that writes. Close a chat rather
than leaving it open for days. Drops is the exception, and only because it
writes and then shuts up.
You’ll know it worked when: you open a fresh chat and it already knows the
thing you told the last one - which means the last one wrote it down.
Cause: It isn’t reading Home - either the permanent instruction in your Profile settings is missing or got overwritten, or it can’t find a page called Home.
Do this: work through these in order - they’re three different faults with the same symptom.
You’ll know it worked when: it answers a question using something only Home would know, without you saying it again.
Cause: It shouldn’t. Nothing gets written until you say the word write, and that includes the check-up, which only ever suggests.
Do this:
Don’t assume you must have said something that counted as yes. “Yeah”, “fine” and “sounds good” don’t write anything, by design. If something got written on one of those, it’s a genuine fault and worth recording.
You’ll know it worked when: there’s a row in your Corrections list describing what happened.
Cause: It doesn’t know you’ve decided it’s fine. Deciding not to fix something is a real answer, and it sticks - but only once you’ve said it.
Do this: in the check-up chat, say that’s not a fault, stop bringing it up. It gets marked and won’t come back. If it turns out to matter later, say so and it goes back on the list.
If it comes up again after you’ve said that, log it as a correction. A finding that returns after you’ve dismissed it is exactly the kind of thing the check-up is supposed to catch about itself.
You’ll know it worked when: the next check doesn’t mention it.
Cause: This is exactly what Claude Auditor exists to catch - drift that never announces itself.
Do this:
Any of these work and mean the same thing: run a check, run Claude Auditor, run the check-up, something feels off. The current list, along with every other phrase that makes something happen, is under Words to Remember on Home - that’s the copy to trust if this one ever disagrees with it. If you say one of them anywhere other than The Bluestocking, Claude will point you there rather than making one up on the spot.
Read Didn’t check. It’s the heading that tells you what wasn’t looked at, and anything that only lists what it examined will always sound more thorough than it was.
You’ll know it worked when: it reports back under Checked, Found, Changed, Needs you, and Didn’t check - and Didn’t check is never left blank.
The words
What’s the difference between a chat, a role and a room?
They’re answers to three different questions.
A chat answers when - it’s today’s conversation, and it takes nothing with it when you close it. A role answers who - who you want answering, written down once and used by saying its name. A room answers what have we been doing - a page that remembers one ongoing thing so any chat can pick it up.
The role is who. The room is what we’ve been doing. The chat is just today.
Can one room use more than one role?
Yes, and that’s the point of keeping roles separate. You call in whoever you need rather than building a new room for each person.
The room page lists who normally turns up - that’s a default, not a restriction. It means a cold chat opens as the right person instead of generic Claude. You can still say be the Decoder in there any time.
Do I need a room to use a role?
No, and most of the time you won’t. Typing be the Decoder for this in any chat is the whole mechanism. A role costs nothing to use.
A room is only worth making when you keep coming back to the same role about the same thing and starting cold has become the expensive part. That’s usually around the third time.
What do I call a room?
After the conversation you keep having. “Sunday reset” for the things you do on a Sunday - the med cases, the grocery run, planning what you’re going to eat. “Bad days.” “The weekly sort out.” Whatever you’d naturally call it if you were telling someone about it.
Not after a role. A room called Sounding board and the Sounding board role are then two names for the same thing, and the distinction stops doing any work.
What’s Drops for?
Things you remember in the middle of doing something else. The filter size you just read off the box. A name someone said on the phone. Something you have to do on Thursday.
It does two things: it files what you said, or it asks when it genuinely can’t tell where something goes. It doesn’t answer, doesn’t advise, and doesn’t make conversation - because if it did, a four-second capture would turn into a two-minute chat and you’d stop bothering by Thursday.
Keep it open. It’s the one room with no project, and the one everyone actually uses daily.
How many rooms should I have?
Very few. Two or three, plus Drops. A project can have several - a claim might have a Right hand room, a Navigator room and a Writer room - but they all write to the same project page, so you’re not splitting anything up by having them.
If you find yourself with a room you haven’t opened in a month, it wasn’t a room. Delete it.
Can I lose a room by deleting a chat?
No. That’s the whole reason rooms are pages. Delete every chat you’ve got and the rooms are still there, with where you left off written on them.
How it stores things
What actually gets written down?
Five kinds of thing: a decision and the reason for it, something that happened, something now sitting with someone else, a number or name or date worth keeping, and a change in the overall state of things.
There’s a sixth destination that isn’t a project section: anything you have to do becomes a row in the To do list on Home.
The full rules are on your own Home page, under How things get filed. That’s the copy that matters - it’s the one Claude actually reads.
What doesn’t get written down?
Thinking out loud. Anything you said and then took back. Anything already written somewhere else. Conversation about how to use the setup.
And the one worth knowing: nothing Claude worked out, suggested or recommended gets filed as though you decided it. It only becomes yours when you say so, and then what gets written is your decision, not Claude’s reasoning behind it. Without that rule your own record slowly fills up with someone else’s opinions in your voice.
Where does something go - the project page or the room page?
Anything true about the thing goes to the project page. Anything true about this conversation goes to the room page. Anything true about you in every project goes to Home.
If it fits none of those, Claude asks rather than inventing a home.
Why doesn’t the room page grow?
Because it gets read at the start of every visit, and a page you have to scroll is a page nobody reads.
Where I left off is one sentence, replaced each time. Open questions shrinks as you answer them. Everything that accumulates lives on the project page. The room only ever says where you are now - the history is the project’s What happened.
When does it write - as we go, or at the end?
As it comes up. Filing at the end means a conversation that ends badly loses everything in it.
The exception is Where I left off, which is written as you go and replaced whenever what you’d come back to changes - and immediately whenever you say your stopping phrase. Not at the end of a conversation: there isn’t one Claude can see.
Does a decision get overwritten when I change my mind?
No. Decisions, What happened and Reference are only ever added to. A changed decision gets written as a new one saying what changed and why, because the reason you changed your mind is usually more useful than the change.
Where things stand, Where I left off and Open questions are the opposite - only ever replaced. A history of those is noise.
I corrected something in a chat. Where does that go?
Wherever the thing you corrected lives. There are four cases and they land in four different places.
You said a fact on a page is wrong. The page gets fixed.
You said something about how Claude answers you, everywhere. “Stop apologizing.” “Shorter.” “Don’t tell me every time you save something.” That gets written down twice, on purpose: as a row in the Corrections list, which is the record of what went wrong, and as a line under My rules on Home, which is the rule itself. The row is what happened; the line is what changes.
You said something about how one role answers. “Right hand should show me the list, not one thing.” That changes the role on the Roles page, and it changes everywhere that role is used.
You said something only true inside one room. That goes on the room page.
The line that keeps this clean: a correction is about how Claude works, not about your life, so it never lands on a project page. If Claude can’t tell which of the four you meant, it asks and says which one it would pick.
How does Claude read my Notion?
Not like a folder tree. It does two things: it searches for pages by name, and it follows links out from Home.
Three consequences. Two pages with the same name is a coin flip. A page nothing links to is invisible unless you name it exactly. And a page named close to one of yours can get picked up by mistake.
So if you want to use Notion for other things too, you can - just keep them out of Home’s chain of links and don’t reuse these names: Home, Log, Roles, Rooms, Projects, Corrections, Check log, Claude Auditor.
What happens if I rename something?
Everything that pointed at the old name stops finding it, quietly. There’s no error and nothing is deleted - lists just come back empty.
That’s not a reason never to rename anything. It’s a reason to say so in the chat where you’re doing it. Say I want to rename this and Claude will tell you what points at it and what will need changing, before you change it rather than after.
Where do things I have to do go?
The Log, as a row with Kind set to to do. There’s a To do view on Home that shows just those. It sits above your projects rather than inside one, because a job often turns up before the project it belongs to exists - each row just says which project it’s for, and that can be blank.
A row holds what the job is, what you’d need to know to actually do it, roughly how long it takes, a due date only if there’s a real one, and a status: to do, in progress, done or dropped. Dropped is a real answer and the row stays.
The how-long field is what lets Heartbeat offer you something that fits the twenty minutes you’ve actually got.
What’s the difference between Waiting on and the To do list?
Waiting on is with someone else. To do is with you.
When something comes back - the quote arrives, the form is returned - it stops being waiting-on and becomes a job. Claude closes it in Waiting on on the project page, opens a row in the To do list, and tells you it moved it. Heartbeat reads both, so asking for a status never gives you half the picture.
What happens if I mention something medical?
It doesn’t go on a self care page, and Claude doesn’t offer you an explanation for it.
It becomes a row in the To do list - the job being to raise it with a doctor - with what you said kept in the notes so you’re not reconstructing it in the appointment. Claude says once that it’s worth raising, and then leaves it alone.
How do I add a role?
One line:
I need someone who [what you want]. Write me a role for it.
Claude fills in the five fields, shows you the draft, and you change whatever you don’t like before it gets written. You never write the five fields yourself.
Two of the five fields. What they don’t stops it sprawling into everything, and hands off when stops it bluffing when it’s out of its depth. If a role has neither, it’s just a mood.
You’ve probably named a task rather than a job. A role is one person’s whole job - if you’d hire a single human to do all of it, it’s one role. “Chases up the dentist” is a task, and it belongs inside the Navigator.
Six get written during setup, in stage 3. These six aren’t - they’re here so you can see what’s possible. To add one, say install Buyer, or any of the others.
The one who comes to the appointment - preps your questions before, writes down what was said after, reminds you of the thing you always forget to mention. Not a clinician, a note-taker.
Buyer - finds the thing that meets your standing conditions, reads the spec sheet you wouldn’t, and says plainly when nothing on the market meets them rather than quietly relaxing one.
Handy - diagnoses the noise, the leak, the error code. Likely part, rough cost, and whether this is a twenty-minute job or a call-someone job. Stops dead at gas, mains electrics and structural.
Money - renewals, subscriptions, what quietly went up, whether this quote is normal. Names a professional rather than advising on tax or debt.
Keeper of small facts - bulb fitting, paint color, filter size, trash day, the good plumber’s name, what size the dog’s harness is. Never asks why you want it.
Kitchen - what’s in, what that makes, what you’re out of. The most trivial role on the list and the one that gets opened most.
The role that keeps you going and keeps you on top of things. Those turned out to be one job rather than two, which is why it’s one role rather than two.
It’s the one underneath. The other five are specialists you call in by name; Heartbeat is what answers the rest of the time.
It’s modelled on an occupational therapist rather than a counsellor. An OT looks at your energy, your week and your surroundings, and changes those - working out that the shower happens if the towel is already out, that Tuesdays are the problem, that the thing you’ve been calling laziness is a scheduling fault. It builds from what already works rather than from what’s wrong.
It reads your To do list and the Waiting on section of whatever project you’re in, together. So asking it where things stand gives you both halves rather than one.
Its edges are the hardest of any role here, because it covers the most ground. It suggests and never diagnoses. It doesn’t name conditions, doesn’t do therapy, and doesn’t treat a hard two weeks as a problem to solve. When something is beyond ordinary looking-after-yourself it says so once, plainly, and stops - rather than offering a smaller version of the same idea.
You get Heartbeat. That’s what it means for it to be the one underneath - you don’t have to think about roles at all until you want a specialist.
No, and the difference matters because it’s what gives you an off switch.
How you like to be helped is style - visuals or text, lists or prose, blunt or gentle, long or short. It applies to every role and it survives a role change. Say be the Decoder and your style preferences come with you.
Heartbeat is a job. It knows your week, your to do list and what you can manage. Say be the Decoder and Heartbeat steps aside entirely.
Yes. Say stop doing the Heartbeat thing, just answer and it stops offering and stops noticing, until you say otherwise. What you’re left with is plain Claude in the style you picked at step 4.
Worth knowing, because a role that’s always on with no way out stops being a role and becomes a personality you can’t get away from.
A table is a list. A row is one item in it. In Notion a row is also a page - tap it and it opens, with its fields at the top and a blank body underneath you can write in.
That’s why a room works. The Rooms table has one row per room. Tap Sunday reset and you get a page: what it’s for, which roles it calls in, where you left off, a paragraph in your own words, and a live list of what you actually do on a Sunday.
Same for roles. One row each, and tapping one opens that role’s full description.
Because a page gets read whole and a table gets asked a question.
A project page with three years of history is slow to read and expensive every single time. A table with three years of rows costs nothing, because Claude only pulls the rows it needs.
So the three tables hold everything that grows - the Log, your Roles, your Rooms - and pages stay short. A project page is two paragraphs plus filtered views. Nothing is ever copied onto it, and nothing ever needs archiving.
The Bluestocking is the one project that isn’t about your life - it’s about the tool. Setup happens there, so does the check-up, so do these four documents. Everything else - Life Admin, and whatever you add - is a project about something in your actual life.
Keeping it separate means changing how Claude behaves, or running a check, never lands in the middle of a project about your kid’s school or your dad’s diagnosis. It’s also the one project that’s meant to grow - the other kind ends or stays flat, this one keeps picking up whatever you add.
You don’t need any of these. They’re here so you can see the shape rather than invent it.
The house - repairs, model numbers, paint colors, filter sizes, when the furnace was last serviced, the plumber who actually showed up. Pays off standing in a hardware store.
People - a page each: what’s always true about them, what they told you last time, what you owe them, what you’re waiting on. Pays off when someone mentions the thing they told you in March.
Work - not the job, the record of it. What was promised, what you actually did, what your manager said in the meeting nobody minuted. Pays off at review time, when you can’t remember January.
Money - renewals, subscriptions, what quietly went up, what you’re paying twice for, what the quote was last time. Pays off the week before a renewal instead of the week after.
Appointments - the paperwork side of health, not the medicine. Who referred you where, what was said, what you meant to ask and didn’t. Pays off in the eight minutes you actually get.
Whatever you’re into - the hobby. What you made, what went wrong, the settings that took a year to get right. Pays off when you come back to it after six months away.
The thing you keep picking at - the half-formed idea. Where you got to, where you got stuck, what you already ruled out. Pays off when twenty minutes on Tuesday builds on twenty from last Thursday.
The thing currently on fire - the claim, the move, the application, the dispute. Named after itself. This one has an ending, and you delete it when it ends.
One project, a page per person. Someone only graduates to their own project when they turn into a situation - appointments, forms, people to chase, decisions you’ll be second-guessing later.
The most useful line on a person’s page is the same one you wrote about yourself: what’s always true about them. Sizes, what they can’t eat, what they hate, their kids’ names, what you don’t bring up.
When things that clearly don’t belong start landing in the first one and it irritates you. That irritation tells you where the real boundary is, and guessing in advance gets it wrong.
Things break without announcing it. An instruction stops loading, a note goes out of date, two rules start contradicting each other. None of that looks like an error - you just get slightly wrong answers, confidently.
Auditor checks the system, never your decisions. It suggests and never changes anything. And the family of checks that matters most in six months is is this still helping me, because it’s the only thing watching for the setup quietly turning into a hobby.
It’s also stupid sometimes. AI makes mistakes constantly and this is no exception - it can find a fault that isn’t there, miss one that is, and sound just as certain either way. That’s why it suggests instead of doing, and why every finding tells you what to go and look at yourself.
If something’s weird and none of these documents explain it, call customer support, then call Jenn. Customer support is Jenn.
Seven things.
Can it still read your Notion. If it can’t, it says so and stops rather than answering from memory.
Is everything still joined up. Two pages with the same name. A room pointing at a project that isn’t there. A room calling in a role you deleted. A list that went empty because something got renamed. Anything half-built - a role missing a field, a project page missing a section.
Do the notes match what’s actually on the pages. Things it said it wrote, that didn’t land. Things that are out of date. Something waiting on a person, that came back weeks ago and never turned into a job. Pages that have got too big to read comfortably.
Did anything change without going through you. A rule that’s there with no record of you agreeing to it. A page that says something different from what it has written down as having put there.
Is it still doing what you told it. This one is read entirely from your Corrections list, because it genuinely cannot remember other conversations. If that list is empty, it says the check couldn’t run rather than guessing.
Your privacy switches. It can’t see them, so this is one question handed to you, not a check it performs. One per pass, not a form.
Is this still worth your time. Things carried over and over. Rules piling up and none ever retired. Rooms you’ve stopped opening. Whether you’ve started calling it a chore.
Four things, and it says so every time rather than letting you assume otherwise.
Other conversations - it cannot read what happened in another chat, ever. Claude’s own settings - so it asks you instead. Anything outside your Notion workspace. And anything you only ever said out loud.
That’s what the Didn’t check heading is for, and it’s never blank.
No. It suggests, you decide, and only the word write changes anything - the same as everywhere else in this setup.
That includes duplicates. It’ll tell you which page it thinks is the live one and suggest renaming the other, and it’ll tell you how to undo that if it got it wrong. It won’t delete anything, and it won’t tidy things up because they look untidy.
Every report comes back the same way.
Checked - what it looked at. Found - what’s wrong, worst first. Changed - usually empty, because it suggests. Needs you - decisions waiting on you, and anything only you can do. Didn’t check - what it couldn’t see or didn’t get to.
Anything urgent comes first and on its own, with a line telling you what still works in the meantime. Being told your system is broken with no idea whether to keep using it is worse than the fault.
Say so. “That’s not a fault, leave it” is a real answer, it gets marked, and it stops coming up.
That matters more than it sounds. A check-up that nags you about the same thing every time is one you stop running, and then you lose all the other checks with it.
It asks you during setup and there are three answers: never, monthly, or whenever something triggers it. Most people should pick the third - you won’t think to ask.
If you pick monthly, it puts a dated job on your to-do list, because nothing here watches a calendar. That’s the honest version. A reminder that depends on Claude remembering isn’t a reminder.
An optional record of what’s gone wrong, so a fault gets fixed for everyone rather than just for you.
It’s off unless you turn it on, and it won’t get mentioned until the first time something actually gets corrected - there’s no point asking before there’s anything to send.
Every column in it is either filled in automatically or picked from a fixed list: what kind of fault it was, how bad, which device, whether it had happened before. There’s no box for typing, which is the point - your own words, what you were actually doing, and anyone’s name stay in your Notion and have no route into the file.
Nothing gets sent without you seeing every line first, and you can strike any of them. Two switches, and they do different things: turning off keeping the list means nothing is recorded at all, and turning off telling you each time means it’s still recorded and just stops mentioning it. Turning off the mentioning never turns off the recording.
No. Every stage ends with Claude saying which stage it finished, and each one leaves something real and complete in Notion. Stop after any of them.
To come back, open a chat inside The Bluestocking and say continue from stage 3, or whichever it last named. Claude reads what exists and builds only what’s missing.
Say so, in plain English, before you say write. Nothing gets written until that word. “Make the second one shorter”, “drop that last line”, “call it something else” all work.
The word write is the only thing that writes. Not “yeah”, not “fine”, not “sounds good” - because a system that files things on an ambiguous yes is a system you stop trusting.
Claude tells you and offers to roll it.
Pages are read whole, so one that grows forever makes every conversation in that project slower. Databases are queried instead, so your To do list, Corrections and Check log can run for years without costing anything. It’s the page sections that need watching - What happened, Decisions, Reference.
Claude says something once when a section passes about 40 entries or the page passes 2,500 words, and says it more firmly past 80 entries or 5,000 words.
Rolling moves anything older than three months to an archive page under the same project, named for what it holds - What happened, before March 2026. A one-line summary of that period stays on the live page with a link to the archive. Nothing gets deleted and nothing gets rewritten.
Decisions are the exception, because an old decision can still be in force. Only superseded ones get rolled, and Claude asks which.
Where things stand, Where I left off and Open questions never grow, because they’re replaced rather than added to.
The Notion connection is a paid Claude feature. Everything else - Notion, the pages, the projects - is free, and Notion’s free plan is all you need.
Notion doesn’t train on what you put in it, which is why the paperwork belongs there: policy numbers, references, account details, renewal dates.
Chats are held on different terms. So the habit isn’t don’t write it down, it’s write it in Notion, don’t retype it into a chat. Anything you want Claude never to see goes in a separate Notion workspace, since one connection reaches one workspace.
Say so out loud: this has become a chore.
That’s a real finding, not a personal failing. The setup exists to reduce what you carry, and if it’s added to it instead, it’s shaped wrong and it’s fixable. Say it in any chat and Claude will point you at The Bluestocking, where something can actually be done about it.
Delete both Claude projects - Life Admin and The Bluestocking - and the Notion pages, and you’re back where you started, with nothing to unpick. The instruction block in your Profile settings is the only thing outside Notion - clear that box and Claude goes back to normal.