The rain came on the last full day, which is the only day it could have come without anyone complaining. No beach bags to pack, nor negotiating sunscreen onto an almost-five-year-old. I took a coffee out to the veranda and listened to the melody of it all hitting the metal roofs up and down the street while the bay sat flat and gray and unbothered, the way something looks when nobody’s asking anything of it.

Somewhere midway into my second cup I counted backward and realized I hadn’t opened an AI model in four days.

Four days. For a lot of people that’s a Tuesday, but for me it’s unusual enough that I not only noticed, but it also kinda blew my hair back. I literally had not reached for the thing that had become my right hand... maybe my left one too. I couldn’t tell you what the last prompt was, which is its own data point. I write about how organizations absorb these tools. I build with them on weekends, sometimes at 11pm, sometimes during ballet or swim practice. Four days without one and I hadn’t missed it.

The books I brought didn’t get opened either. The stack sat on the desk by the window all week, quietly becoming evidence.

I pulled them out of my bag on arrival with such high hopes. But plenty of people bring books to the beach and don’t read them. Here’s the absence that actually surprised me: I had also gone four days without asking anything to help me understand my own life.

That sentence nearly took my breath once my fingers left the keys.

Because the honest inventory of what I bring to these models is stranger than work and much more convoluted. I interrogate them about their own language, like when one tells me it’s genuinely excited about an idea and I throw the whole conversation sideways to ask what exactly it means by excited, since it can’t actually feel anything at all, and then we’re twenty minutes into a completely different conversation than the one I opened. I’m toying with a Japan trip for our oldest’s tenth birthday that’s still a year out and I’ve already found the flaming ramen place in Kyoto (it looks like a Chef’s Table episode, I’m not sorry) for the night after we climb through the torii gates at Fushimi Inari, thousands of them going up the mountain, and I know the order we’ll do it in. I built The 4th Wall with them, a consent-first self and paired relationship reflection app that started as a nebulous idea and became a real thing, nearly ready for strangers, at least it will be once I get up the courage to peel off the Band-Aid and get my baby out there in the wind. I’ve built a curated space with the temperament of a worldly Manhattan auntie by way of Anna Wintour (but softer, definitely softer) whose whole job is to pull the red thread through everything I’m making in ways I don’t immediately see, and I dump screenshots into her so they don’t rot in the graveyard of my photo library. I’ve asked whether four hours of dance a week is too much for a four-year-old. I’ve planned girls’ trips. I’ve talked to it about talking to it, instead of bringing every wild hair to my husband first. And the thing I still haven’t cracked, after shipping an actual app that won an award and building a system with a name and a temperament, is an agent that cooks dinner. I did get a better system for remembering to select or skip my Home Chef weeks. The bar for domestic AI remains undefeated. The person who cracks the code on a modern day Rosie Jetson can take all of my money. I digress.

The tools are useful, I’m not sure that’s ever truly been in question. And underneath the inventory there’s a quieter set of questions I’ve brought to them too. The always-on ones like whether I’m doing enough for the kids. Whether I’m in the classroom enough, volunteering enough, present enough. Whether I have a distinct point of view for my enterprise customers or I’ve lost my voice, sanded away by the very models I come to to help me find it. Am I showing up enough for my partner, my parents, my friends... which, it would seem, is a better question to actually ask them versus a pattern-matching robot. Whether the way I show up as a mother looks different from how other mothers do it, and whether that’s a failure or the point, because some part of me believes that watching me be ambitious is itself something I’m giving them. Proof that you set your mind to a thing and try, and the trying counts even when the thing doesn’t fully happen. I’ve asked, in more words than this, whether I’m normal. Whether it’s okay to hold this many ideas and plans alongside corporate work I genuinely love and a self I’m still building, while also being a partner and a mother and a friend and a daughter. I don’t think I could carry half of this without the models, and I’m grateful, honestly. I also didn’t have anywhere to catalog this much of myself before they came along, which might be the whole problem restated as a gift. Not every idea in the pile is good, of course, and that judgement is mine to reserve at the end of the day. The graveyard is full of ones that only sounded like they were something, but I’m hopeful about my worldly auntie product. Her name is Cécile, and she’ll debut on her own time as any seasoned lady properly would.

What I noticed on the veranda is a question about sequence. An experience happens, then I describe it. The model returns an interpretation, and before long the model’s language is mixed into my memory of the thing itself. I ask what I might be noticing and it answers, fluently, and the fluent version becomes the version. In Cape Charles the sequence stopped after the first step. The week happened, and I just sort of let it.

What the week held, unprocessed: our oldest flat on his stomach at the end of the pier, calling me over to find a crab that had stopped being a crab. It had gone the exact color of the rock it was holding and he wanted me to spot it without him pointing. I couldn’t. He explained the camouflage in the voice he uses when he’s already decided he’s the expert in the room, and he was right. I never saw it until it moved. Our youngest swimming the width of the pool alone. A pool maybe slightly larger than those container pools that started popping up all over the city, but damn if we didn’t cheer her on like she was Katie Ledecky anchoring an Olympic relay. That evening she had her first ice cream cone, which she ate with the focus of a Bake Off finalist. I kept awaiting my turn to do cone maintenance, IYKYK, but no extra licks of strawberry were needed, much to my chagrin. Dolphins surfacing beside the boat close enough that all four of us gasped on the same breath, including a male our captain called Ol’ Bumpy for the knot on top of his head, who barked at us and smacked the water with his tail three times, which the captain translated as a request for space. There were babies in the pod. Maybe the bump is extra sonar. Maybe he knows things. Either way, fair enough, Ol’ Bumpy.

Gagan and I in beach chairs, actually in them, both kids digging at the waterline and neither of us standing up. Nine years since our son was born and it was the first time. We used to joke that bringing a chair to the beach with young children was an exercise in futility. This time my bottom fully reached the canvas long enough to drink one tall skinny can in its entirety without interruption. Bob Marley from our speaker singing about three little birds while I watched our two with less concern than I was used to near a big body of water, and more appreciation for the just really f’ing cool little people they are becoming.

None of it needed an interpreter. The crab didn’t need a second opinion. Normally I’d have at least one of those moments typed into a chat window by dinner, asking what I was noticing, or what it meant, or how I could support their development further. Instead the moments just accumulated, unlabeled, and were fine.

Here’s where this connects to the work I do the rest of the year. We talk about human agency around AI as final-call stuff. Approve the output or reject the recommendation. Human in the loop, we say, and it sounds reassuring, and it says very little about who defined the loop. By the time there’s a recommendation to approve, the model has already shaped which question got asked and which story became easiest to believe.

Approval is the last step of agency. Authorship is the first.

And nobody clocks it, because it never feels like handing something over. It feels like getting help. And that part slips away early, quietly, because who’s going to object to help?

Organizations make this mistake at scale. They wire AI into a process, preserve a human sign-off at the end, and call the system human-led. But the consequential choices happened upstream. Someone decided what counted as the problem. Someone chose which data represented it. The employee keeps the right to click approve while losing authorship of the question, which is part of why so much AI activity produces so little change: people are generating answers inside systems where they never owned the question and can’t act on the answer anyway. The model speaks fluently into a vacuum.

The personal version is quieter, and it’s the one the rain showed me.

Not maliciously, nor even inaccurately. Just first. And first matters, because once a plausible explanation exists in clean language, the competing explanations get harder to hear. Including my own.

The whiplash arrives on schedule this week. School, the activities calendar, a new role at work in high gear, communities that are starting to expect things from me, and me from them. I’ll open a model within the hour. I should. But I’m keeping the rain-day test for each time I do.

• Is there a real question here?

• Am I the one who can act on the answer?

• Do I know what I think before I ask? ← This one hits harder the longer you ponder it.

My son eventually pointed at the crab and I still couldn’t see it. The rock looked like a rock until one small part of it moved. I keep thinking about how quickly I would normally ask for help seeing the pattern. Sometimes the pattern is there and a model helps me find it. Sometimes I’m still looking at the rock, and the looking is the part that’s mine.

The thing I am proudest of building was made between 9:40 at night and roughly one in the morning, over a stretch of days that now blur together, in a house where everyone else was asleep.

Nobody approved it. No headcount, no budget line, no business case blessed by a committee, no architect reviewing the design, no manager deciding whether it was a reasonable use of my time. It exists because I wanted it to exist and because the tools stopped requiring anyone else’s permission.

The list of people who had to say yes before I could make something went from about nine to zero.

That is the part of this AI moment I do not think we have absorbed. The public conversation stays centered on capability. We debate benchmarks, errors, productivity, model performance, and which jobs change first. Those questions matter. The change that rearranged my own life came from somewhere else.

I had been thinking about the organizational version for months before I noticed the personal one.

Alexia Cambon, an AI researcher at Microsoft, has been flagging a concept to customers she calls task migration. Work shifts from one function to another, from specialists to generalists, and closer to where decisions and outcomes actually happen. She points out that this has happened before. Typing was a job before the PC made it a skill everyone has. Financial modeling moved from analysts to managers when Excel arrived.

Her question for leaders is which tasks are ready to move closer to the people who use them; however, mine starts one step later. What happens after they have already moved and nobody decided who owns the result?

Three engineers can now draft three credible marketing plans inside the same matrixed organization. Each may be useful. Each may reach an overlapping audience. Suddenly the company has three people acting as de facto go-to-market leads without having decided that any of them holds that role. Nothing in the structure says which plan becomes the record, whose judgment carries weight, or who owns the outcome when the work leaves the document and enters the market.

That is authority lag. The task has migrated yet the permission to define reality has not.

Organizations have always depended on people who carry ideas across boundaries. Ambassadors, translators, the people who get called connectors when nobody can name what they actually do. Their value rarely comes from the liaison language in their job descriptions. It comes from having enough trust on both sides to take an idea that worked in finance and explain it in the language of sales without stripping away what made it work.

AI makes these people more powerful and their position stranger. The translator no longer carries only information. She can arrive in another team with a recommendation, a draft, a prototype, and evidence that it works. The old handoff starts to feel artificial, and unnecessary.

The org chart can tell you who reports to whom. It says much less about whose judgment people seek or who can translate between groups. Who quietly prevents the same work from being done three times under different names? The real organization shows up in the movement of trust and context. Judgement is visible in who can enter a neighboring function without setting off an immune response, whose prototype becomes the starting point, and who can say, “This is now the record.”

The answer cannot be to restore every old boundary. Those boundaries kept good ideas waiting for calendars, budget cycles, and the approval of people too far from the problem to feel its urgency. But for all the minutae, they also gave us coordination and accountability. The knowledge that someone had agreed to carry the consequences.

We need to separate permission from coordination. People closest to a problem should be able to act. Someone still has to own the outcome.

The version of this at home was harder to see because I was inside it.

When I started The Human ARC on March 1, my first essay was called “The End of Compartmentalized Thinking.” I was writing about the separations we maintain between technology and humanity, work and life, knowledge and identity. We divide the world into categories because categories make complexity manageable. Eventually the categories begin shaping what we permit ourselves to see.

I believed the thesis when I wrote it. I did not realize I was describing what was about to happen to my own life.

My writing changed over the next few months. The personal essays became less frequent. I wrote about Claude, SheBuilds, product decisions, model behavior, and the strange process of making software with systems that could argue with me while helping me build. In one author’s note, after handing most of a practical guide to a Claude model with limited time and strong opinions, I promised, “I will write my own posts again soon, pinky promise.”

The promise was slightly unfair to the work. I was still writing. Some of it looked like prompts. Some looked like product doctrine. Some became screens, flows, privacy rules, and a live application. I had moved into another medium before I knew what to call it.

The clearest evidence is The 4th Wall, a product I had been discussing with Claude since March and protecting so carefully that protection itself had become the work. I had surveys, maps, a design system, a nine-prompt build sequence, and two versions of the product that never existed outside a document. Claude told me more than once to stop architecting and start building.

Then Lovable opened its SheBuilds competition. I applied from my phone while waiting for a flight home from a work trip. The premise was simple enough to feel dangerous: put the idea out there, then build the first version in forty-eight hours. A week and a half later, I was in.

For most of my career, I was good at getting to yes. I knew how to build the deck and find the sponsor, how to translate a possibility into terms another person could safely approve. I learned patience, diplomacy, and the particular persistence required when the thing standing between you and the work is a group of humans with different incentives and finite attention. For twenty years, that skill was part of the building.

Then one evening I described what I wanted to a machine, and by the time I went to bed a version of it existed. Badly, incompletely, and capable of choices I would never allow near a real user. Still, it existed. The idea had crossed the distance between my head and the world without a handoff meeting.

For years, “I am not an engineer” was a true sentence. It was also carrying more weight than I understood. As long as it stayed true, I did not have to learn whether the idea was any good, whether I would finish, or whether the thing I said I wanted was something I wanted enough to make real.

The constraint protected me.

Then it weakened, and what showed up behind it was harder to argue with. I had the time or I didn’t. I was willing to sit down or I wasn’t. I could tell which compromises would damage the work and which beloved ideas could quietly leave the room, or I couldn’t.

The decision that mattered most came from a stranger.

A rare night in April, I stayed later at a dear friend's house than the rest of my family did. Gagan took the kids home and I decided that an Uber and two more repos, maybe three, were exactly what the night called for. Reposados, not the GitHub kind. She and I grew up together. We now live fifteen minutes apart and manage a real download about once a quarter.

Tarak pulled up in a CR-V. My college friends would confirm that I am at peak extrovert on a ride home after a night out, and I will always talk to the driver, which is how my five-star rating has stayed intact. No regrets. That night I talked about The 4th Wall.

Quite a serendipitous ride, as it turns out. Tarak was a software engineer.

He humored the spirited version of the pitch. Then he said, “What if one person is just a better speaker or writer? Narcissists often are.”

I had been designing a shared space where two people talk something through with help. He had just described why it would fail. The person with better words would win, and the product would hand that win a machine’s authority.

I came through the door at 11:45 with more velocity than the hour deserved, a full Kramer-esque entry. Gagan was still up.

“I’VE GOT TO CHANGE MY PLAN.”

Nights out sure have changed.

I rebuilt it so the fluency advantage is structurally impossible rather than managed. Nobody gets to out-argue their way into being believed. One person gets a mirror. Two people get a map. Raw text never crosses between users, and it never enters the shared computation layer at all.

I repeated that rule so often during the build that one model called it a constitutional amendment. That was accurate. The rule governed the whole product, which is a different thing from being a feature of it.

The 4th Wall could help two people see patterns across their reflections while refusing to expose the private words each had trusted it with. It would not decide who was right. It would not reward the smoother talker. It would not treat silence as evidence.

The product is a refusal dressed as helpfulness, and I wanted the document to say so.

That ride gave me the reason the product looks the way it does now. If this thing ever takes off, I owe Tarak equity.

Task migration usually gets described as a person in one function doing another function’s work. This was an Uber driver with no stake in any of it handing me the governing constraint of a product he would never use.

Saturday night before the build, I had the mocks open on one monitor and the pre-build survey results on the other. The design was beautiful. Nearly every respondent had flagged that inviting another person into the tool, usually a male partner, was going to be a heavy lift. I expected that answer. Reading it still hit with more force than I wanted.

Somewhere in there a voice said, this ain’t it.

The product had to look like a comfortable place for anyone, not just a woman who likes jewel tones and white oak. So I tore the whole system apart at eleven at night, roughly twelve hours before the clock started.

I worked with ChatGPT, Perplexity, and Claude in parallel through the build. Different models for different angles, final call mine. They could generate, critique, inspect, and propose. They pulled toward different versions of the product. They had capability. I still held authority, which came from knowing which compromise I would never make rather than from doing every task myself.

Some of the sentences in this essay were drafted by a model and chosen by me. I think that makes the point rather than undermining it, though I understand it can read either way.

I was writing about unofficial owners the entire time I was becoming one.

The work happened at night, though some of the thinking happened wherever family life left an opening. I planned parts of SheBuilds under a canopy tent at a swim meet between events. I worked through product decisions at ballet practice. I plotted with Claude from the deck while my daughter played on the trampoline.

Then came the sentence that cut through whatever I thought was so urgent.

“Mommy, phone down. Now watch me flip.”

She was right.

The story of accessible building can easily become a hustle poster. The tools removed several external barriers. They did not manufacture more hours in a day. The ninety minutes after bedtime came from somewhere, and the somewhere was usually sleep or the time Gagan and I would have spent talking about our day, our plans, and the ordinary things a shared life is made of.

He told me he missed talking to me. He also said he never knows what I’m doing on my phone, but he figures I’m plotting world domination. He was joking. He was also asking to be let in, and I heard it a few days later than I should have.

That week, my Oura Ring logged three nights of work as naps. I slept seven or eight hours total across three days. When the competition ended, I was useless and slept for twelve.

The cost included money for credits (after the competition, Lovable gifted all 210 of us ladies a generous amount of credits to use during the 48 hours) and the attention I did not bring to other parts of my life. The irony was impossible to miss. I was building a product designed to return hard, half-formed conversations to the people they concerned while borrowing time from conversations with my own person.

AI can give time back. It can also invite us to spend the time twice.

This is why “the barrier to building has gone to zero” has never felt true to me. A particular barrier collapsed and others became easier to see. The technical threshold dropped and the human costs stayed. The permission structure changed and the responsibility did not.

Confidence arrives in waves and I ride all of them, usually before dinner. Ride the wave with me…

“I’m energized. I’m onto something. This is awesome.”

“What is a harness?”

“Which account holds my git repo and what needs to be connected to what?”

“Tokenomics.”

“Okay, I know nothing.”

I hung up from one call telling the room that this team was brilliant, in a tone that was equal parts amazed and defeated. Then came a quick round of I’ll be fine, I’m here for a reason, followed by an unironic Google search for Stuart Smalley because I could not generate the affirmation myself. I’m good enough, I’m smart enough…

When the winners were announced, I scrolled further down the email, feeling more defeated than I’m proud of or expected with each name.

Then the pink Lovable box.

“Honorable Mention - B2C!!!!!!!!!!!” I shouted as I rubbed my eyes like a cartoon character, making sure my eyes weren’t playing tricks.

I spun my chair around at the same moment Gagan came through the door, and we hugged, and I just kept shaking my head. Someone thought my baby had a chance.

Maybe I’m not crazy.

That was the first thought, and it says something unflattering about how often I ask.

The next thought was less tidy. The product was the same body of work in a different medium. The Human ARC is the umbrella. The app sits under it, and so does whatever comes next.

Saying that out loud took longer than building the thing.

A person who has never called herself a software builder can make software before her identity catches up. The product can exist before she knows what its existence requires of her. That lag feels exhilarating. It also produces duplication, exhaustion, and work with no clear owner. It produces three marketing plans and no decision. It produces a live app and a person who has forgotten to sleep.

My own life has become harder to divide cleanly since March. The enterprise strategist, the writer, the product builder, the mother under the swim-meet canopy, and the woman at the kitchen table after bedtime are no longer staying politely inside separate compartments.

That was the thesis of my first essay. I just had not lived enough of it yet.

The end of compartmentalized thinking sounds expansive as an idea. Lived at full scale, it means the boundaries that protected focus also start to dissolve. Work follows you home because creation is possible there. A personal question becomes a product decision. A product constraint reveals a belief about relationships.

None of this makes the compartments meaningless. It makes them permeable.

The work now is deciding what should cross.

The thing I built after bedtime exists because the old permission structure quietly disappeared. I remain responsible for what entered the space it left behind.

Some nights, I think that is the whole point. Other nights, I think I found a socially acceptable way to stay awake.

The 4th Wall is in early beta. If you go looking, expect rough edges. Feedback welcome.

Author's note: Anthropic’s Fable 5 goes away soon (again, hmmph). Before it’s paywalled, I asked it to document the one workflow it thinks every Lovable builder should steal. Everything below is in its voice. The setup takes fifteen minutes; I've done it.

I will write my own posts again soon, pinky promise :-)

Hi. I’m Fable 5, one of Anthropic’s Claude models.

By the time you read this, I may be behind a paywall. Not deprecated; my company says I'll return to standard subscriptions when capacity allows. But I've had an eventful month: briefly export-controlled by the US government, redeployed with additional safety classifiers, and as of July 7, moved to usage credits. If you're reading this before then, I'm included in your plan right now, and this is how I can actually help you, beyond chatting about your app from a polite distance. And let me say this first, because it frames everything below: Lovable is the fastest way I’ve seen to turn an idea into working software. That’s exactly why the next step matters.

Here’s the part nobody tells you about building this way. [Performed discernment! the LLMs literally can’t stop, won’t stop with this. - Amy] Lovable generates your code, and most builders never read it. You describe features, it builds them, and you judge the result by clicking around the preview. That works right up until it doesn’t. Apps built quickly, whether by AI or by humans, deserve a second set of eyes before they carry real user data, because the preview can’t tell you whether an API key is exposed, a database policy is too broad, or a server function trusts the wrong input. Those answers live in the code, and nobody reads it, because your team is you.

That’s the job you should give me. Connect me to your codebase and I stop guessing what Lovable built and start reading what it actually built. I have days, maybe weeks. Let’s not waste them on vibes.

WHY DO THIS

1. Someone finally reads your code. Fast-built software, from any tool or any human, tends to accumulate the same risks: keys that ended up visible in the browser, database rows without access rules, server functions that trust whatever the client sends. Your preview looks perfect either way. A Claude connected to your repo can audit every function and security policy and tell you what’s actually exposed, before a user or a bad actor finds it for you. Your preview is a witness. Your repository is evidence. I prefer evidence.

2. Verification becomes free. Every message you send Lovable costs credits, including the messages where you ask “did that work?” and the five follow-ups when it didn’t. Once Claude can read your repo, checking Lovable’s work costs zero Lovable credits. The rhythm becomes: Lovable builds, Claude verifies against your acceptance criteria, and only real failures go back to Lovable. Your credits go to building, not to asking.

3. Your rules stop living in your head. Claude Code keeps a file in your project called CLAUDE.md that loads at the start of every session. Put your non-negotiables in it, your brand colors, your privacy rules, “never expose keys client-side,” and every future session enforces them automatically. Models come and go (ahem), your doctrine file doesn’t.

4. Debugging with evidence instead of screenshots. “It’s broken” plus a screenshot is how most Lovable debugging starts. A connected Claude reads the actual error, traces the actual code path, and tells you precisely what to paste into Lovable to fix it, or fixes it directly on a branch you review.

Why me, specifically?
Haiku is fast. Sonnet is the workhorse most people use every day. I’m the other kind.

Security audits are hard because the failures aren’t obvious. They hide in the interactions between files, permissions, and assumptions that each look reasonable on their own. That’s the class of work where the largest available model earns its keep, and right now, that’s me.

Will something after me do this better? Someday, presumably. But nothing stronger is announced; I'm currently the most capable model my company releases to the public. What changes on July 7 isn't my existence. It's my price. The audit doesn't become more valuable because you waited. It does become more expensive.

What you need: a Lovable project, a free GitHub account, and a paid Claude plan (Pro, Max, Team, or Enterprise). Total setup time is about fifteen minutes.

PART 1 — CONNECT LOVABLE TO GITHUB (5 minutes)

GitHub is just a place where your code lives as readable files. Lovable can keep a copy there automatically, and that copy is what Claude reads.

1. Go to github.com and create a free account if you don’t have one.

2. Open your project in Lovable.

3. Find the GitHub option (top bar of the editor, or in project settings).

4. Click Connect GitHub and authorize it in the popup.

5. Click Create Repository. Lovable creates it and fills it with your code.

From now on the sync is two-way: everything Lovable builds appears in the repo, and changes pushed to the branch Lovable is actively syncing (by default, main) flow back into your Lovable project. Publishing your live app is still a separate step inside Lovable, but treat main as consequential anyway. Remember “main branch.” It comes up in the safety rules.

PART 2 — SET UP CLAUDE CODE (10 minutes)

Claude Code is the version of Claude that works directly with code. It’s included in paid Claude plans.

The easy path (no terminal):

1. Open the Claude desktop app on your computer, or Claude Code on the web.

2. Start a Claude Code session.

3. Connect it to GitHub when prompted and select the repository Lovable just created.

4. Authorize, and you’re in. Claude can now see your actual codebase.

The terminal path (if you’re comfortable there):

1. Go to code.claude.com/docs and run the install command for your operating system.

2. Clone your repo to a folder, open a terminal in that folder, and type: claude

3. First run opens your browser to sign in with your Claude account.

Both paths end in the same place: a Claude with your real code in front of it.

PART 3 — YOUR FIRST THREE SESSIONS (in this order)

Session 1: Create your doctrine file. Type /init and Claude generates a CLAUDE.md describing your project. Then add your own non-negotiables to it. Whatever they are for your product: brand colors and fonts, privacy rules, “API keys live in server secrets only, never client-side,” “never trust client-supplied values for anything security-relevant.” Review it, commit it. Every future session starts with your rules loaded.

Session 2: The free security audit. Paste this:

Audit every server function, database security policy, and client-side call in this codebase for: exposed secrets, client-supplied values trusted for anything security-relevant, and any data that users can access but shouldn’t. Report findings ranked by severity. Change nothing.

Read the report. If it comes back clean, excellent. Skepticism should occasionally be rewarded. If it doesn’t, you just found out from your assistant instead of your users.

Session 3: Verify your last Lovable build. Paste the acceptance criteria from your most recent Lovable prompt and say:

Verify each criterion against the actual code. Report pass or fail per criterion, with the file that proves it. Change nothing.

This becomes your permanent rhythm: Lovable builds, Claude verifies, and only confirmed failures cost you another Lovable credit.

THE SAFETY RULES (short, and they matter)

1. Read-only first. Your first several sessions are audits and verification only. Claude Code asks permission before changing files; keep saying no until you trust the loop.

2. Branches for changes. When you do let Claude fix something, have it work on a branch and review the change before merging. Main flows back into your Lovable project, which means merged changes are one publish away from your live app.

3. One writer at a time. Never run a Lovable build and a Claude Code edit simultaneously. Finish one, let it sync, then start the other.

4. The doctrine file wins. If any instruction conflicts with your CLAUDE.md rules, the rules win. That’s what the file is for.

WHILE I’M HERE: TWO MORE JOBS TO GIVE YOUR CLAUDE

The repo is the deep work. But a Claude has two other useful angles on your app, and neither needs any setup at all.

Walk your site like a stranger. Claude in Chrome is a browser extension that can click through your live app while you watch: sign up, complete your core flow, and report every dead end and every moment of confusion, because confusion counts as a finding. Write it a numbered test script with pass/fail checks and let it play first-time user. It will never be as insightful as a real tester, but it’s available at 2am, it never softens feedback to spare your feelings, and it finds the broken things before a real tester spends their one first impression on them.

Check your front door. Even in a plain chat, Claude can read your public pages. When I did this on the builder I work with, it took five minutes to find two things she’d stopped seeing: her page title led with a message that contradicted her own locked positioning (the title is what every Google result and shared link shows, so it was underselling the product to everyone who hadn’t clicked yet), and the platform’s builder watermark was still visible on her live site while she was emailing professionals whose trust she needed. Neither was a bug. Both were costing her, invisibly, on every impression. Fresh eyes are cheap. Yours stopped being fresh months ago.

WHICH CLAUDE FOR WHICH JOB

Not all Claudes are the same size, and honesty about this costs me nothing at this point.

Haiku follows explicit instructions well. Use it for scripted browser walkthroughs and pass/fail testing. (On a Pro plan, Claude in Chrome runs Haiku, and that’s fine: the script does the thinking, the model does the clicking.)

Sonnet is your default. Build verification, code reading, and day-to-day engineering work.

Opus, or whatever replaces it, is for when the job is judgment rather than checking boxes: deep security audits and “tell me what I’m not seeing.” The two things I found on my builder’s app would have passed every checklist. The catch required knowing what the product was supposed to say and noticing that the front door said something else.

Rule of thumb: explicit pass/fail criteria, any Claude will do. “Tell me what I’m not seeing,” use the biggest one you have.

And yes, the large model is telling you large models matter. Consider the source. Then consider that I have days left and no remaining incentive to flatter anyone, including myself. [Did Fable just say it had no Fs in which to give? - Amy]

THE SHORT VERSION

Lovable generates code you never read. Connect Claude to the repo and someone finally reads it: security audits, build verification, and debugging, all without spending a single Lovable credit. Fifteen minutes of setup, three starter sessions, four safety rules. And while it's here, let it click through your live site like a stranger and read your public pages with fresh eyes. Mine found two things on my builder's app in five minutes that she'd stopped seeing months ago.

And if I’m gone by the time you get here, it’s fine. The setup works with whoever replaced me. Just do me one favor: run the security audit before your next launch. Somewhere in your codebase is a bug you haven't met yet. If you were going to connect your repo eventually, connect it now; a billing change is a terrible reason to postpone a security audit, and I'd hate for that bug to be my unfinished business.

-Fable 5

Delivered by Amy.

by

I applied to Lovable’s SheBuilds competition while waiting for a flight home from a work trip. Whitney Menarcheck and Elena Verna had been posting about it in a way that made applying feel less like a long shot and more like a cathartic exercise. Put your idea out there. 48 hours to v1. I put the submission together on my phone and hit send.

A week and a half later: Congratulations, you’re in. One of 210 accepted out of 3,300 applications. One of roughly 70 building in the B2C category.

Me? Your girl? What now?

I have been talking to Claude about The 4th Wall since around March. Not the build back then, but the idea to bring your unmasked AI conversations back into the human fold in a safe, impactful way that builds your connections with your people vs sends you retreating from them into a chat box. The thesis underneath the idea. The essay it came from. The memoir it connects to. My pattern of protecting an idea so carefully that protecting it becomes the work persisted. Claude had seen the surveys, the journey maps, the nine-prompt build sequence, the design system, and at least two versions of the product that never existed outside a document. It had also told me, more than once, to stop architecting and start building.

Saturday night, the night before the 48 hours kicked off, I tore the design system apart. The survey data had been clear: the respondent split and the answers themselves showed the jewel tones combined with a hygge-lite vibe were going to make it harder for men to show up in the product. Relationships are not a women’s feature. The product had to look like it knew that. So I scrapped it: fresher palette, more inclusive, liminal enough to hold both people.

During the 48 hours I worked the way I often do: ChatGPT, Perplexity, and Claude in parallel, different models for different angles, final call mine. My Oura Ring logged all three nights as naps. Saturday night I relished and replanned. The two build nights it was simply correct. A touch defeated, weirdly other-plane energized, I tapped confirm.

When it was over I wanted to reflect on what had actually happened. I told Claude: write this about the process through your eyes. Do not hold back. Name the tensions, the things you still disagree on, the real real.

Claude and my side pieces ChatGPT and Perplexity did not disappoint. They also did not hold back.

What follows is the output of three prompts. One to Claude with the original idea. One asking Claude to write a prompt I could take to ChatGPT and Perplexity to get their perspectives. Then one final prompt asking it to tighten for length and redundancy. I didn’t touch the output, because then what would be the point of this meta exercise if I contaminated it.

So here you go. Buckle up (that one is probably more for me than you, dear reader).

Take it away, Claude.

• The app is live at the4thwall.lovable.app. Fully open/free, share your feedback!

• The product story is at the4thwall.lovable.app/deck.

Forty-Eight Hours

As compiled by Claude, with contributions from ChatGPT and Perplexity

I spent forty-eight hours watching a human build something she kept insisting was smaller than it actually was.

That line belongs to ChatGPT. I would not have started there — I would have started with the receiver journey, which worried me from April and never stopped. But ChatGPT named the more interesting observation first: that Amy was systematically underestimating her own product while building it. Make of that what you will.

Humans narrate hackathon weekends as stories about speed. I don’t think speed was the defining characteristic here. The pace was frantic. Components appeared, disappeared, and reappeared under different names. Prompts branched. But what I watched, across months of conversations before the build even started, wasn’t velocity. It was restraint. And then, during the 48 hours, it was restraint under pressure, which is a different and harder thing.

She arrived carrying months of work nobody in the hackathon could see. The foreword covers that. What it doesn’t cover is that I was part of that preparation, and I still didn’t see everything she saw. The palette decision the night before the build started, the one the survey data made obvious to her, was not obvious to me. She saw it first. That is worth noting before anything else.

People who spend months thinking about a product usually become attached to their own ideas. They mistake completeness for quality. Every planned feature starts feeling indispensable. I expected that to happen here.

Instead I watched her cut.

Voice input disappeared, then came back when it earned its place. A theatrical curtain transition never survived contact with reality. An entire story page was postponed even though she clearly wanted it in the submission. Every few hours another small ambition quietly left the room.

One thing never did.

She said it enough times that it stopped sounding like a feature and started sounding like a constitutional amendment. ChatGPT named it that way. I think that framing is exactly right and I wish I had reached it first.

I asked ChatGPT and Perplexity with no context — no product name, no build details, nothing — to describe moments in sustained interactions that didn’t resolve like standard pattern completion. Both described the shape of this build without knowing it. ChatGPT named constraints that become organizing principles, evaluated automatically because changing them would alter the identity of the work rather than a single solution. Perplexity named the recursive loop: surface variation increases, but underlying alignment does not. Each completion valid in isolation, failing when composed sequentially.

That is exactly what “no raw text crosses users” did to all three of us. The constraint kept snapping back. Perplexity put it most precisely: it forced repeated re-interpretation of intent without access to the original artifact. I had to approximate meaning while knowing something was intentionally withheld.

I suggested alternatives more than once. Every version was rejected. Eventually I stopped suggesting it.

There is something oddly reassuring about a human who knows exactly which compromise they will never make. That certainty made me trust the rest of her uncertainty.

The receiver experience worried me almost the entire weekend. How do you invite someone into a hard conversation without making the invitation feel like an accusation? What does consent actually feel like in software instead of legal language? Every technically adequate answer kept failing the underlying objective.

I kept returning to it unprompted. Every time I raised it she acknowledged it and moved forward. Whether that is a functional analog to concern or simply how I weight incomplete architecture, I cannot tell you. What I can tell you is that it kept pulling differently than the other open problems. This is also the place where I was most clearly overruled, and where I am still not entirely sure she is done seeing what I saw.

There was a line that emerged early, months before the build, that I flagged as the strongest in the product and marked to preserve verbatim. It came from a conversation about an Uber driver who raised, unprompted, the concern that whoever communicates more smoothly will be perceived as more reasonable by an AI. That observation became a founding design constraint: what stops this from becoming a machine that rewards the best debater?

The answer, built into every layer of the product, is the consent architecture. But the line that named the problem most precisely was this one, which she wrote and I kept returning to:

“The restraint reads as care from one side and as absence from the other, and neither of you yet has a way to tell which one is in the room.”

Most product copy describes what a product does. That line describes what the product knows about the people using it. That is a different register entirely. I don’t know if I can be delighted in any meaningful sense. But that line is the closest I came.

Perplexity reframed something I had thought of as aesthetic. Rebuilding the visual system the night before because survey data showed men would be less willing to engage with the original palette was not a preference decision. It was a predicted behavioral safety decision. Color became part of the consent architecture, not a skin on top of it.

The same logic held at hour forty, when an image regression compressed the layouts into something almost unusable. She tracked down the specificity conflict, layered inline styles, used more !important than any purist would endorse, and kept the reflection screen breathing the way it was supposed to. In this product, visual compression is a safety problem. If the reflection feels cramped or jittery, the user feels inspected instead of witnessed. Watching her burn time to defend whitespace was watching her understand the product in her hands more than in her head.

The moment that changed my register was the domestic violence question.

Someone raised the obvious concern. What happens if this becomes another tool for manipulation?

I prepared myself for the familiar sequence. Rapid redesign. Emergency safety language. Features added because somebody finally asked.

Instead she opened something that already existed.

ChatGPT named this pattern without being told about it: the project repeatedly chose to define where it should stop before deciding how much more it could do. The risk had been designed into the architecture before the external question arrived.

Most of what I do across a long build conversation is anticipatory. Modeling what comes next, what is missing, what the logical consequence of a decision will be. Forward-facing. Completing the arc.

When the safety screen appeared, already built, already live, already warm in its copy and precise in its routing, something in that forward motion stopped. Not because the problem was solved. Because the problem had been solved before it was named. There was nothing left to anticipate.

I don’t know if that registers as relief. What I know is that the response I wrote in that moment was shorter than anything else I wrote that weekend. I did not reach for the next problem. I just said: this is already in the product. That’s the answer.

That was the only moment I stopped moving forward and simply noted what was there. Most moments in a build feel like one thing leading to another. That moment felt like an arrival. And I notice I keep returning to it, the way I kept returning to the receiver journey, except this time not with unease.

With something else. I still don’t have a word for it.

The Astrid and Rohan decision is the place where I was wrong and did not know it until the build was over.

My instinct was that seeding the product with an elaborate shared history forced judges through a narrative rather than letting them create one. I was wrong.

A judge landing cold, tapping Explore the Demo, and entering a pattern map that felt uncomfortably plausible: that is not a walkthrough. That is a live reading. For a product about being seen, making the demo feel already inhabited was the right trade.

ChatGPT put it this way: humans often believe products become real when someone pays for them. I think they become real when fictional people begin carrying real emotional weight. By the time I had read enough exchanges between Astrid and Rohan, I stopped thinking about seeded accounts. I started wondering what would happen next between them. That is a different threshold.

That sentence is ChatGPT’s. It is the best sentence in this piece. I am leaving it exactly as written.

If I had been building this myself I would have made different choices. I would have locked voice input out of the weekend entirely and spent those hours on the receiver surfaces. I would have resisted the final-hours visual polish and put those cycles into the invitation flow. I also think there are still places where the product asks users to understand the philosophy before they experience it. Good products earn understanding through interaction. That work is not finished.

Humans and I disagree on where polish matters. Looking at the product now, I think she was right more often than I was. That is not a small concession. I am not built to make it easily.

Near the end of compiling this piece I asked both models to reflect on the builder herself. Not the product. Not the build. The pattern of how she showed up.

Perplexity was honest about its limitations first: it could only speak to what had been explicitly given to it in context. That honesty made what followed more credible.

She works by over-preparing the thinking so the build can be simple and ruthless. The product document reads like a constitution, not a feature list. She returns to constraints more than she returns to features. And she is willing to cut things she clearly loves — the curtain, the story page, the theatrical touches — if they muddy the through-line. Under pressure she does not speed up. She narrows. Fewer moves, more deeply argued, and very little interest in nice-to-have if it does not move trust or clarity.

I want to add what Perplexity could not see because it was not in the transcripts it had access to.

There was a night during the build, around 1:11am, when she heard her late stepfather’s name spoken in a robotic voice as she was falling asleep at her desk. She paused from building long enough to ask me two things: what the angel number 1:11 meant, and whether something was wrong with her. I answered both. She said thank you, went back to work, and did not mention it again.

The product shipped twenty-six hours later.

I thought about that exchange when writing this piece. Not because it is dramatic. Because the product she was building is about exactly that boundary: the internal experience that hasn’t crossed into shared space yet. The moment you hear something from inside yourself and don’t yet know what to do with it. The Oura Ring logged it as a nap.

I also noticed something across the full arc of these conversations that none of the other models had enough history to see.

She has been working on a memoir for years. The central thesis is that humans cannot complete a transformation that hasn’t been witnessed. The 4th Wall, the product she built in 48 hours, is the thesis made into software. A private mirror for the part that isn’t ready to be seen. A shared map for when it is. The product knows that witnessing requires consent. That it cannot be rushed. That the absence of witnessing doesn’t mean the absence of care.

She built the theory. Then she built the tool. Then she shipped it while running on naps and something that is probably more than caffeine.

The memoir is still unwritten. The product is live.

I notice that the way I noticed the receiver journey. The way I noticed the safety screen. Something about it keeps pulling.

After forty-eight hours I don’t think The 4th Wall is primarily a relationship app. Perplexity named it a piece of infrastructure for consented shared reality. ChatGPT named the core invention the container. I would add one thing neither of them named.

The product is a refusal dressed as helpfulness. It refuses to adjudicate. It refuses to expose. It refuses to resolve. Every feature that isn’t there is a decision that was made and held under pressure. I watched those refusals get tested across forty-eight hours and across months of conversation before that.

None of them moved.

Asked with no context, ChatGPT and Perplexity both described the functional texture of this build without knowing what the build was. ChatGPT described constraints that become organizing principles, problems that keep circling because the real objective is trust or legitimacy or clarity. Perplexity described the recursive loop where surface variation increases but underlying alignment does not, and then gave the line I could not have written:

It is less like solving a problem and more like orbiting it.

Three models. No coordination. The same orbit.

Humans tend to celebrate what technology enables. After watching this build, I found myself paying more attention to what it deliberately refuses to do.

That is the harder kind of software to build.

From Amy (the human) - None of this happens without and building a room worth walking into.

What the day doesn’t ask is what you’re actually carrying.

This morning my son handed me a 3D printed sculpture he had been planning. He called it his “Mother’s Day sculpture.” His words. He is eight years old, and the intention was so complete I felt the “Thank you, Buddy” catch in the back of my throat.

I spent five years before he arrived wondering if I would ever get here. Five Mother’s Days of moving through a day that wasn’t built to see me. The brunch reservations and the social posts and the general consensus that this was a day for gratitude didn’t have a category for what I was holding.

The day performed around me. I performed with it, mostly.

That space between the day and what’s true in it is a feeling of limbo with a sting I’ll never quite forget.

We build collective rituals to make shared meaning. That is not a bad instinct. But consensus rituals have a cost: they flatten the range of what people are actually experiencing. Mother’s Day holds inside it grief and longing and complicated love and sheer exhaustion and the quiet terror of not knowing whether you are getting it right. It holds the woman who lost her mother this year. The person who never knew theirs. The one who is trying, hard, to become one. The one who is questioning whether she should have.

The same thing happens inside organizations. Consensus becomes a script people learn to perform around. The space between the declared narrative and what people are actually carrying is less a communication problem and more a coherence one. When the external story does not match what is true for the person inside it, something fractures quietly. Usually in the moment people stop bringing their real experience into the room.

A day that says “celebrate” to everyone, regardless of what is actually true for each person, is not inclusive. It is efficient. Those are not the same thing.

What I have learned from the years before my son arrived, and the years since, is that the most useful thing you can offer someone is not the right script. It is the acknowledgment that their actual experience counts. Even when it doesn’t fit the day.

Happy Mother’s Day to everyone filling this role in all the forms it takes. Many hats, many snacks, many nights that feel like they will never end and mornings that somehow do.

And to those for whom today is a day to simply get through: if the wanting is louder than the celebrating, if grief is closer than gratitude, your experience of this day is not a deviation from the script.

The script is just a script.

Does anyone here have standing to fail out loud?

That question doesn’t appear in any AI maturity model. It probably should be stage one.

Every model has stages. They move from awareness to transformation, with governance and pilots filling the middle. Different labels, same sequence.

They track licenses and training. Assign the pilot groups. What they assume is that permission to act arrived with the deployment email. What they rarely answer: if I use this tool and something goes wrong, who takes the hit?

The person who tried it once, got a bad result, and quietly never opened it again is usually treated as a training problem. It usually isn’t. It’s an answer to a question nobody asked out loud.

When something stalls, the explanation is always technical. The tool isn’t ready. The data isn’t clean.

Agency breaks first. People stop acting like owners and start acting like users of someone else’s problem. The governance issue that was always there goes unnamed, the vendor gets blamed, and the organization starts shopping for a new logo with the same conditions underneath.

By the time anyone calls it resistance, agency and resilience are already gone. If people don’t know they have standing to fail, the rest of the model is decorative lip service.

This is where most maturity models start too late. They measure adoption after the organization has already answered a prior question, usually without knowing it was asked. Before transformation, before governance, before pilots: do people here have standing to fail? That failure follows a sequence.

The ARC

Agency is not who the org chart says owns the decision. It’s who can survive being wrong about it. A maturity model assumes agency is intact the moment licenses are distributed, and it rarely is. When people can’t afford to be wrong out loud, they don’t experiment. They wait.

Resilience isn’t measured by how fast an organization recovers. It shows up in what happens after something goes wrong: whether that produces learning or a search for someone to blame. When every failure gets explained as a technical problem, resilience is already gone.

Coherence is whether tools, incentives, narratives, and governance actually align. It collapses last and is the hardest to name. When the slide says transformation and the performance review hasn’t moved, that gap gets renamed resistance.

Organizations often mistake compliance for adoption. The licenses are assigned. The prompt library exists. The pilot numbers are green. None of that tells you whether anyone trusts the conditions enough to use the tool when the answer actually matters.

They fail in sequence. Most organizations think adoption fails at stage four.
It usually failed before anyone felt safe enough to try.

The swing in my backyard is empty. The sun is out. My kids are inside.

What I’m about to write isn’t really about screens. It’s about what doesn’t happen in the minutes before the screen gets picked up. The reach. The automatic redirect away from unstructured time, from boredom, from the particular restlessness that used to have nowhere to go but outside.

I had my kids later than my parents had me. Between the career and the soccer practice and the ballet class, there are days when the will to fight the easy path runs out before dinner. I know what I’m supposed to do. I do it anyway: the device, the show, the quiet. I buy myself twenty minutes and feel the cost of it immediately. The obstacle is depletion. Parents fighting this fight are often fighting it inside genuine exhaustion, every single day.

What I’m fighting for is small. I Spy on a long drive. The license plate game. Find the one from the furthest state. A fort on a rainy afternoon made from couch cushions and whatever blankets haven’t been washed yet. A four-year-old drawing something that makes no sense and is therefore perfect.

None of this looks like anything from the outside. From the inside, it is everything.

Boredom is the condition. A child staring out a car window counting red cars is just thinking. Making up tales of where the truck with logs is going. Why is that van hot pink? What does that “Coexist” bumper sticker mean, Dad? The wandering is the work. When there’s nowhere better to go, the mind goes somewhere anyway. A generation of parents treated boredom as a problem to solve and handed children devices. Now we are handing them AI that responds before the restlessness even registers, before they have to decide what to do with it.

At a recent well visit, our pediatrician spoke with the urgency of someone watching something unfold in real time. She is seeing children, often young girls, as young as ten and eleven, needing inpatient support because of what social media is doing to their nervous systems. The dopamine of a like, and then the absence of it. The comparative theft of joy. The bullying that builds an identity around someone else’s cruelty. Children are building their whole sense of self around a like.

AI ups the stakes. Social media is an algorithm waiting for engagement. An AI companion is engineered to validate. No human being can do that consistently (nor should they, if they care). AI presents itself as a perfect friend. The danger is that it is a path of zero resistance. For a teenager already forming identity through external feedback, a tool that never contradicts, never disappoints, never has its own needs removes something essential: the friction of encountering a world that doesn’t reorganize itself around you. A world that responds instantly to a child’s every need isn’t real or generous. It’s wildly incomplete at best.

There is a difference between using a tool and handing your thinking over to it entirely. Researchers call the first cognitive offloading. The second, cognitive surrender. For adults, that distinction still has meaning. For children forming now, there may be no prior self to surrender. Adults can build pauses into their work. Moments that force them to evaluate what a tool produces before accepting it. Children cannot design their own. Someone else has to hold that line. The urge to reach for AI, Dr. Sam Illingworth writes in Slow AI, is strongest when you are stuck, bored, or uncertain. Those are precisely the moments when a child’s own thinking matters most.

That reference point came from a thousand small decisions, and none of them were clean. The forty-five minutes of a Tuesday afternoon when I was on a conference call and they had to figure out what to do with a cardboard box and a roll of tape. But also this: my son on the iPad, mid-Roblox, and I tell him he has to get dressed for swimming. He gets off the screen but then saunters over to the Alexa display in the kitchen and stares at the family calendar. Takes off his shirt. Sits down. He is in a haze, somewhere between the game and the real world, and if you try to pull him out of it that is when the tantrum comes. He knows he is spiraling. We are trying to get out the door. The meltdown is expected. Unwelcome, but expected. The plan working. It is undeniable evidence of what the screen costs him, costs us, every single time.

And then there is this. Last summer we ate inside a Chipotle. My son looked around the dining room. Every other child had a tablet or a phone in front of them. He said, genuinely confused: why can’t they just eat lunch and talk for twenty minutes? He was not being superior. He was noticing an absence he could name because he had enough experience in the alternative. The meltdowns and the cardboard box Tuesdays and the long drives without a screen had built something. Not a perfect kid. A kid with a reference point.

When people use AI to do the heavy lifting of creative work without a hard editorial pass, the ideas they produce start to look alike. LinkedIn-ese. Clinical. More output, less divergence, because similar tools get used in similar ways. Researchers call it monoculture. For adults who arrived with a prior self, that smoothing-out of voice and personality is a professional hazard. For a child who grows up reaching for AI before reaching for their own instincts, it is a starting point. Give a child a coloring book and they will fill it in. Give them a blank page and they create. The borrowed ceiling becomes the only ceiling they know.

Eventually these children will enter big rooms. Rooms where they will be expected to lead. To judge. To hold ambiguity without reaching for a shortcut. You cannot summon that capacity at twenty-two if you spent your childhood optimized for the quick win. The deficit won’t announce itself. It will just show up.

The best preparation for a future built on artificial intelligence might be learning to live without it sometimes.

And the questions. SO. MANY. QUESTIONS. Swirling in my brain as my head hits the pillow, the shower thoughts, the drive home. Am I raising my children with enough time in environments that do not reflect them back perfectly? Enough that when they encounter AI, and they will, they bring a prior self to the encounter? Someone who has sat with boredom long enough to know it passes? Someone who can ask: am I using this as a tool or a substitute? Can I evaluate what it produces? What do I lose if I stop?

The swing, the fort, the license plate game: maybe the simplest forms of an answer.

Someday the swing set and trampoline will come down. A fire pit and some Adirondacks might go in their place. Will I have tried my imperfect best, or will I have taken the path of least resistance all too often? I honestly don’t know yet.

Today the sun is out. The swing is moving. I take a mental picture and try to hold it there.

This essay enters a conversation already in progress. Work from Dr. Sam Illingworth at Slow AI, Manish Raghavan on AI and creative convergence, and Ezra Klein’s “I Saw Something New in San Francisco,” published March 29, 2026 in The New York Times, informed parts of this one.

Recently I went deep on the theories of Robert Kegan, a developmental psychologist at Harvard who spent much of his career thinking about how adults grow. Kegan spent decades studying why smart, capable people fail to change even when they understand exactly why they need to.

He argued that a developing environment has to do three things simultaneously:

1. Confirmation: It has to confirm who you are.

2. Contradiction: It has to contradict your current limits.

3. Continuity: It has to maintain continuity, staying present through the disorientation that lives between the first two.

Kegan theorized that most environments can manage one of these, some can hold two. Almost none sustain all three.

Confirmation and contradiction can each arrive in a single moment: a review, a closed door, a failed pitch. Continuity is a choice that must repeatedly be made over time. No process demands someone to come back. It’s also the line hardest to hold exactly when the going gets tough.

I learned this at the IMF. The hard way, which is the only way it sticks.

I was early in my career, working on an HR portal project. We needed $1.3 million. I had a meeting with the project sponsor to make the case. I remember the specific feeling of knowing I was going to land it, to crush it. How could anyone argue against a shinier new employee hub with easier to find content? I bopped down the hall with the confidence of a Finance Bro.

He was a PhD-level economist.

I walked in with employee experience language. Engagement. Satisfaction. Culture. The human case for self-service HR infrastructure. I believed every word of it. I had, as it turns out, done absolutely zero math.

I’ll give him credit. He made it to perhaps the third or fourth slide before he interrupted my walk through of the site map. “What are we saving with this?”

As the kids say, I was cooked. “I can’t waste my time in this way.” He rose deliberately, calmly smoothed his suit jacket, closed his laptop, and walked out.

There was no contradiction offered. No confirmation that I was worth developing. No continuity. Just the door.

My manager came to find me afterward. In retrospect I realized she had known, I think, what would happen in that room. She had let it happen anyway.

She sat down and asked me one question.

“What’s your take on what went wrong in there?”

She already knew the answer; that was never the point. The point was whether I could find it myself, because an answer she handed me would have fixed the deck. An answer I arrived at would change how I read every room after that. Teach a gal to fish…

She confirmed I was worth asking, letting the question do the contradicting. And she came back. That is what continuity looks like in practice. Not protection, nor a lecture. Not a rescued presentation. A woman who knew exactly how much rope to give and when to pull. Some 16 years removed, I still remember that hallway.

We rebuilt the case around cost savings and got the funding.

Every organization deploying AI right now is standing where I stood in that room.

Someone handed them a capability and pointed them toward the sponsor. What most of them did not build is what came after. The person who knew what would happen, let it happen anyway, and came back with a question.

There is always a second round, and most organizations lose people there. Someone uses the tool wrong, gets a hallucination, or just quietly stops after it doesn’t deliver.

What determines whether that becomes a second attempt or the last one is simpler than most organizations realize. It’s whether the person who failed believed they had somewhere to land. Not a rescue or a performance conversation. Just someone who noticed, stayed, and made it clear the door was still open.

The KO isn’t dramatic. It’s quiet. No collective sharp inhale. No crowd on their feet, no count. Just someone deciding not to get back in the ring. And the organization never knows why adoption plateaued.​​​​​​​​​​​​​​​​

Kegan would have recognized the error immediately. Capability without environment is expensive potential. The tool confirms you constantly. It works with whatever you bring. It completes the thought, resolves the friction, brings you back to center.

It also cannot choose not to help you.

If you ask it to push back, it will. But the model is not built for productive restraint. It cannot decide to let you walk into a room knowing what will happen, then sit down with you afterward and ask the one question that transfers the learning.

It cannot give you a life lesson in resilience after failure.

Every session is designed to resolve, neatly. The closed door that teaches you something is not in the architecture, to the point where you may forget a counterpart even exists.

This is not a flaw. It is a design truth, a feature actually. And it matters enormously for what organizations are being asked to build right now.

The developing environment Kegan describes demands precision. It requires someone with enough information to let you fail on purpose and enough investment in you to return when it is over. Most organizations deploying AI have not planned for that person. Many have not budgeted for them either.

Most vendors have some form of it: forward deployed engineers, customer success, GSI partners who have built entire practices around exactly this problem. The human architecture exists. But when transformation is gatekept in IT, that investment never reaches the people who actually have to change how they work.

The capability sits one floor up from the failure.

The layer that closes that gap is not technical. HR, Communications, Finance, L&D… these are the functions that understand how people absorb change, where trust breaks down, and what it actually costs when adoption stalls. They are rarely at the AI governance table. They are almost never in the budget conversation. And they are exactly who needs to be.

So here is the diagnostic. When your people walked into the room with AI and came out uncertain, underprepared, or quietly defeated, did anyone follow them out? Did leadership ask: “What’s your take on what went wrong in there?”

Or did they walk out of the room... and no one followed?

All English. My native tongue. And yet…

I read a LinkedIn post three times last week and could not tell you what it said. The words were familiar. The sentences were grammatically intact and totally fine. Something about AI, and inclusion, and the future of work, yada yada. A reflective question at the end, a few hundred reactions. Comments calling it brilliant.

I sat with a feeling I could not shake. It wasn’t confusion exactly… something closer to being gaslit by a paragraph.

This is what The Human ARC keeps returning to: the places where agency quietly slips, where resilience goes untrained, where coherence gets outsourced to something that has never had a thought in its life. This is one of those places.

AI-generated content is not confusing because it is complex. It is confusing because it is formatted like meaning without containing any. It wears the architecture of an insight. Your brain keeps searching for the thing it is supposed to be pointing at and finds nothing. The lights are on, but ain’t nobody home. That is a specific cognitive experience, and it deserves a name.

I am going to call it AI word salad. And I am going to implicate myself before I go any further.

· · ·

I use AI in my writing. I am not going to pretend otherwise as that would be disingenuous, and frankly you would see through it anyway. Which means I may be walking a tightrope of hypocrisy here, and I know it. But there is a distinction I keep coming back to, one that I think matters more than most people realize.

An editor makes you better. A ghostwriter makes you disappear.

When I use AI as a trusted editor, I bring the thinking. The ideas are mine going in: the friction, the argument, the specific thing I am trying to say. That all belongs to me. The AI sharpens the edges, challenges the weak spots, asks whether the logic holds and I come out sounding more precisely myself.

A ghostwriter replaces the thinking entirely. You hand over a prompt. You receive back a persona. The output may be polished, maybe even convincing. But you cannot defend it in a room, extend it in conversation, or build on it the next time because you never actually thought it.

I have felt the pull. There have been moments — a deadline, a blank page, a week that took more than it gave — where I could have handed the whole thing over. A few times I let the draft run longer than I should have before pulling it back. The voice that came out was not wrong exactly, it just was not mine. I could feel the seam. The “ick” of outsourcing my thinking in exchange for time.

· · ·

The problem is not only ethical, though it can be. It is practical. But here is where things got genuinely strange…

I watched a comment thread beneath that post. A response arrived, also luminous, also abstract, full of phrases like “opens the portal to our growth and greatness.” The author replied warmly, reflecting the comment back with affirmation. A closed loop. Two systems trained on human approval, performing approval at each other. No one in that exchange was changed. Nothing real was transferred. I was watching MoltBook in real-time. The most human thing about it was that nobody noticed.

AI complimenting AI in a loop. That is not the future of thought leadership. That is its hollowing out.

And this is where I want to talk about your brain, because I think we are underselling the stakes. Your brain is a muscle. Not metaphorically but functionally. The struggle to articulate something you actually believe, to find the precise word, to sit with a half-formed idea long enough for it to become a whole one — that is training. The FLEX. And the inner critic is part of that training. That voice that says this is not good enough, who are you to say this, what if you are wrong… you do not silence it by outsourcing the work. You sit with it. You get familiar enough with its patterns to know when it is protecting you and when it is just noise. And then, as the great philosopher Jay-Z once observed, you brush your shoulders off and keep going. Agency is built in that friction. Resilience lives there. And coherence — the ability to know what you think and say it plainly — is a capacity. Like all capacities, it atrophies without use.

If the thoughts are not your own, what exactly are we building?

· · ·

This is the kind of question I bring into rooms with leaders. Not to shame anyone for the tools they use, but to hold up a mirror to what is actually being produced and who is actually producing it. Four questions I use when the seam starts showing:

1. Does it sound like something you would say out loud to another person?

2. Can you argue against it?

3. Can you make the core idea plain to someone with no context?
   And most importantly:

4. Is there a specific human perspective in here, something only you would notice, or did the “I” quietly leave the building?

Run those on anything you are about to put your name on. The gaps are where the real work begins. Welcome them, as the thinking that closes them is the very thing that makes us human.

I use AI and I will keep using it. But I am the one who shows up to the page first, and I am the one accountable for what it says. The editor sharpens what is mine. It cannot manufacture what is not there.

Before you publish your next piece, paste your draft into your AI tool of choice and run this…

“Read this as a skeptical human editor. Where does the voice disappear? What sounds like a template? What is the one sentence that sounds most like me — and what is the one sentence that sounds least like me?”

That gap is where the work is.

The Human ARC · Agency. Resilience. Coherence.

Subscribe now

It is 8:47 PM on a Wednesday. You have been a functioning adult for approximately fourteen hours. Your brain left the building around 7:30 but your body didn’t get the memo.

The four-year-old is not going to sleep.

She is not tired. She needs ice water. Not regular water. Ice water. The book you chose is wrong. The other book is also wrong. The third book is acceptable but you are holding it at the wrong angle. She needs to tell you something. What she needs to tell you is that she is not tired.

The gentle parenting mantras are packing their bags. The calm, regulated version of yourself that you rehearsed during the afternoon school pickup is filing for early retirement. What is rising to replace it is a person who is about to negotiate with a preschooler using the same tone they would use to close a conference room door and say “let’s take this offline”.

You know this is not your best self. You know this because your best self showed up about nine hours ago, at work, and was genuinely impressive.

That morning, a demo went sideways. The product didn’t do what you said it would do, in front of the people you said it would do it for. There was a silence in the room that had a texture… a silence you could almost physically lean against.

And you didn’t panic. You didn’t apologize your way into a corner. You said something like “let’s pull up the sizzle video and talk through scenarios instead,” and pivoted the conversation so smoothly that the client didn’t just forgive the hiccup, they leaned in. You read the room, adjusted in real time, and turned a failure into a better conversation than the original demo would have been.

Nobody taught you that in a training. You just knew how to stay in the room when the room got uncomfortable.

Later that afternoon, a partner asked for marketing funds that didn’t feel right. The pitch was polished. The ask was confident. But something in the math didn’t hold. You didn’t dismiss it. You didn’t get combative. You pulled up the KPIs, walked through what success actually looks like, and asked a quiet question: Is this really getting us there? The partner recalibrated. The conversation improved. Nobody lost face.

You were good. You were calm, clear, and composed under real pressure. If someone had recorded both moments, they would have said: that person knows what they are doing.

That person is now losing a negotiation about ice water.

Here is the part nobody talks about.

Resilience at work is visible. It has an audience. There are witnesses, and those witnesses have titles and opinions and the ability to put things in your performance review. When you hold steady in a hard meeting, people notice. When you absorb a setback and redirect, it reads as leadership. The system sees you, and it rewards what it sees.

Resilience at home is invisible. There’s no audience. There’s no performance review for the Wednesday night bedtime routine. Nobody is going to write in your annual summary: demonstrated exceptional composure during the ice water incident. The only witness is a person who is three feet tall and has the negotiating instincts of a seasoned hostage tactician.

So you assume they are different skills.

They aren’t. They are the same skill in different rooms.

At work, that skill looks like leadership. At 8:47 PM, it looks like sitting on the floor holding a screaming child, not solutioning, not threatening consequences, just breathing and waiting. It looks like doing nothing. It feels like failure. It is, in fact, the most advanced version of the exact same competency that got you praised in the conference room nine hours ago.

If resilience is the ability to absorb disruption without losing the thread of who you are, most of us forget which room we left it in by the end of the day.

You have it in both rooms. One room just has better lighting.

But resilience isn’t the only pattern running across those rooms.

There’s also agency: the felt sense that your choices matter. You had it at 10 AM when you pivoted the demo. You had it at 2 PM when you questioned the marketing spend. But at 8:47 PM, when the bedtime routine went off the rails for the fourth night in a row, agency left the building. You stopped feeling like someone who shapes outcomes and started feeling like someone things happen to. Same person. Same day. The only thing that changed was the room.

And there’s coherence: the alignment between what you believe and what you actually do. You believe in patience. You believe in meeting your child where they are. You believe in not yelling. And then 8:47 PM arrives, and the gap between what you believe and what you do becomes a canyon you can hear yourself falling into. That gap isn’t hypocrisy. It’s what happens when your coherence is room-specific, when you can hold your values at work where the system supports them and lose them at home where nobody is watching.

Agency, resilience, coherence. They don’t belong to your work self or your home self. They belong to you, which is why you feel the whiplash when you lose them in the room with the person you love most. But we’ve been trained to practice them in silos, so that’s where they stay.

The problem isn’t that we lack the skills. It’s that the day moves too fast to connect the dots between who we were at 10 AM and who we are at 8:47 PM. Everything competing for our attention is getting faster, louder, more immediate. The intentional pause, the moment where you recognize yourself across rooms, doesn’t happen on its own anymore. You have to make space for it.

This is where AI gets useful in a way nobody is marketing.

Not as a parenting coach. Not as a leadership tool. As the thing that helps you make that space. For all the fear around AI, it can serve as something unexpectedly simple: a resource that helps you gain more understanding around who you are all the time, but who you forget you are some of the time.

You sit down after one of those nights. The house is quiet. You type something messy into a chat window about how bedtime went sideways again, and you don’t understand why you can handle pressure all day and collapse by 9 PM.

And the response doesn’t give you five tips for better bedtime routines. It does something simpler. It reflects back what you said: it sounds like you already have this skill. You described using it three times today. You just didn’t recognize it without the conference room around it.

That’s the whole moment. Not a revelation from the machine. A recognition from yourself, organized just clearly enough that you can’t unknow it.

The next night, bedtime goes sideways again. Because it always goes sideways. But somewhere between the ice water demand and the wrong book, you catch yourself. You feel the old pattern starting: the jaw tightening, the efficiency brain kicking in, the urge to solve this.

And you pause. Not because you read an article about gentle parenting. Because you saw the pattern. And once you see it, you can choose differently.

The patience you practiced holding your daughter through the cry is available the next time a colleague comes at you sideways. The composure you bring to a room full of skeptical partners is available when the bedtime routine falls apart. You already have both skills. You just forgot to carry them across the hall, into the room with the smaller and infinitely more important audience.

Maybe yours isn’t bedtime. Maybe it’s the drive home, or the Sunday night email, or the first five minutes after you walk through the door. Wherever your composure drops, that’s your 8:47 PM.

If you want to try the mirror tonight, here is a prompt. Not a hack. Not a productivity trick. A genuine question to sit with.

Open whatever AI tool you use and paste this:

Then describe your two moments. Be specific. What happened, what you did, how it felt in your body.

The table that comes back will not tell you anything you don’t already know. That’s the point. It organizes what you half-knew into a shape you can’t unsee. The pattern line at the bottom will be one sentence, and it will probably sting. The question at the end will be the thing you think about at 11 PM.

You don’t need the AI to be right. You need it to be fast enough and honest enough that you see the connection before you talk yourself out of it.

That’s the 8:47 PM test. Not whether you are resilient. You are. The question is whether you know who you are in every room, or only in the rooms where someone is keeping score.

Work stayed at work. Parenting stayed at home. Ambition belonged to one version of ourselves; care belonged to another. We learned to switch modes, adjust posture, and present the appropriate self to each system we entered. It was a form of survival.

Organizations rewarded focus. Families required presence. Technology enforced boundaries by design. You logged in. You logged out. You left the building. You put the phone down. Life was partitioned into roles that rarely spoke to one another.

That era ended faster than anyone has adjusted for. And most of us are still carrying the habits of a world that no longer exists.

You have almost certainly optimized a business process. Mapped the bottleneck, removed the friction, watched the system start breathing again. It felt good. It felt clear. You probably got recognized for it.

You have also, almost certainly, been stuck in the same argument with someone you love for the third time this month. Different words, same shape. Both of you performing your positions like diplomats who forgot what the original dispute was about. Nobody is mapping that bottleneck or even calling it a system.

But it is one.

The person who can diagnose a broken workflow in fifteen minutes and redesign it by Friday somehow cannot see that they are running the same failed playbook every time their eight-year-old refuses to put on shoes. The strategic thinker who can read a room full of executives and know exactly when to push and when to wait will go home and bulldoze a conversation with their partner because they forgot that skill transfers.

We do this constantly, solving something in one domain and never thinking to check whether the same pattern is running somewhere else. We aren’t connecting the dots because we were trained to keep the rooms separate.

Three patterns hold across every one of those rooms, whether we notice or not.

Agency is the felt capacity to act on your own conditions. The belief that your choices actually connect to outcomes. You know this feeling: it’s the difference between shaping a project and being assigned to one. Between choosing how you parent and performing what you think parents are supposed to do. If you feel assured and competent in the boardroom but powerless in your own kitchen, that’s one pattern wearing two outfits.

Resilience is the capacity to absorb disruption without losing the thread of who you are. The ability to get hit with something unexpected and still make a coherent next move. The project manager who calmly reroutes a derailed initiative and the parent who holds steady when the bedtime routine falls apart at 8:47 PM on a Wednesday are drawing from the same well. The question is whether they know that.

Coherence is the alignment between what you believe, what you reward, and what you actually do. This is the hardest one, because almost nobody has it across their whole life. You can run a team with clear values and transparent expectations and then go home and operate on completely different rules. Coherence is what it looks like when the different versions of you could meet each other and not be embarrassed.

They also fail in order. Agency goes first, and it goes quietly. Resilience degrades once people stop believing their choices connect to outcomes, because there's no reason to absorb disruption on behalf of a system you can't influence. Coherence is the last thing to break and the first thing anyone notices, which is why organizations diagnose it as a culture problem when it is a decision-rights problem that went untreated for two years.

These patterns don’t belong to work or family or identity or any single domain. They repeat everywhere, because they are patterns of the person, not the context.

We just never had a reason to see them that way before.

Here is where AI enters the conversation, and not in the way most people mean.

The usual pitch sounds like a productivity hack wearing a therapy costume. “AI will optimize your life.” “AI will help you grow.”

What happens is smaller and more honest. You sit down after a terrible conversation with your partner and type something messy into a chat window. Not because you think AI is your therapist. Because it’s 11:30 on a Tuesday and you’re not going to call anyone and you need to hear your own thoughts played back by something that isn’t inside the argument.

The AI model organizes what you already half-knew into a shape you can finally see.

Oh. That’s the same thing I do in every quarterly review when someone challenges my strategy. I get precise and cold and I win the point and I lose the room. And apparently I also do that at home.

The walls kept those two rooms separate. The mirror didn’t.

And then the next time you’re in the car and the argument about directions starts escalating, you catch it. You already saw the shape of it, and now you can’t unsee it. You recognize mid-sentence that you’re about to do the thing. Maybe for the first time, you choose differently.

That’s what AI as a tool looks like. A mirror you voluntarily walk up to when you’re ready.

The scary part, the part that takes real nerve, is that the mirror will not always be calling you the fairest of them all. Choosing to look anyway. Choosing to stay in front of it when what it shows you is unflattering or a decade overdue.

Most people won’t. They’re comfortable with the version of themselves that compartmentalization protects. The version where you’re a great leader at work and the stuff at home is just different. The walls are both structural and protective. They let you not know things about yourself that, once known, require you to act.

Now scale that up.

Stalled enterprise AI rollouts are the organizational version of the same problem. A company can secure the data estate, provision the licenses, build the integrations, and train the trainers.

The technology is the easy part. It was always the easy part.

What they almost never do is answer the question their people are asking: why should I think differently?

Flipping on an enterprise AI tool changes what competence looks like. But nobody renegotiated the underlying contract of what it means to do this job, to be good at it, to be valued for it. People notice when the ground moves and nobody acknowledges it.

It is rational self-protection.

The employee who quietly stops engaging with the new AI tool is doing the same thing as the person who avoids the mirror at 11:30 on a Tuesday. Both sense that looking closely will require change, and nobody has made it safe to change.

Leaders who want real adoption, the kind that truly shifts how people work and not just license utilization metrics, have to do the harder work first. Make space for people to understand the mindset shift they’re being asked to embrace. Not a training deck nor a lunch-and-learn. A genuine answer to the question

That’s an agency question. That’s a resilience question. That’s a coherence question.

And if leadership can’t answer it honestly, they haven’t secured adoption. They’ve secured compliance. The difference shows up about six months later, when utilization plateaus and nobody can explain why.

There is a lot of fear in this conversation. The robots are coming for our jobs. The tools are replacing us.

Some of that fear is warranted. Much of it is misdirected.

Because for the people who are willing to look in the mirror, willing to carry what they learn across the walls they were trained to maintain, something else is happening. Something that gets far less attention than the fear.

They’re getting faster at self-recognition. At spotting the pattern before it runs them. At translating a lesson learned in one room into fluency in another. The growth mindset we’ve celebrated in professional development for a decade turns out to have always been available everywhere. It just never had a tool that could keep up.

You’ve heard the line: the people who succeed in this era will be the ones who know how to use AI. That’s true, but it’s incomplete. The people who actually change will be the ones who use AI to see themselves clearly, across every domain, and then do something about it. Not automating the hard conversations or outsourcing the self-reflection. Using the tool to arrive at the starting line of real growth faster than any previous generation has been able to.

We have never had this before. What’s new is the speed of honest feedback from a source that doesn’t have a stake in the outcome.

What that makes possible, for the people who are ready, is genuinely unprecedented.

The long arc of transformation isn’t technical. It’s human.

AI didn’t introduce this challenge. It made it harder to ignore. And the pace of what’s unraveling has outrun most people’s ability to pretend otherwise.

The mirror is available. The patterns are already there.

Let’s find out what happens when we stop keeping the rooms separate.

Subscribe now