Hey y’all,

There’s a research finding from a journal called European Psychiatry that I want to put in front of you because I think it changes the conversation a lot of women in their thirties and forties are having with themselves alone.

A team of researchers trained a machine-learning model on routine medical records. The model learned to identify adults who were likely to be diagnosed with ADHD — months before the formal diagnosis happened. The model didn’t read a personal essay or interview the patient; it looked at the pattern of what showed up across a series of 15-minute visits over a period of years.

I want to be careful about what this means and what it doesn’t mean. The AI is a predictive tool, not a diagnostic tool. It flags people whose records match the pattern of those who eventually got diagnosed. Whether the diagnosis is correct still requires a real assessment by a real clinician. But I think the flag itself is the biggest news.

Subscribe now

Here’s why I think this matters to you.

Adult ADHD in women is one of the most under-diagnosed things in medicine (70-80% of women diagnosed with ADHD didn’t receive a diagnosis until adulthood). And the symptoms in adult women just happen to look like the things women are told they should feel anyway: exhausted, never quite caught up, managing a million things… and even the stereotype of a flighty/ditzy or agitated woman.

Those symptoms (fatigued, brain fog, scattered, agitated, forgetful, overwhelmed, etc.) are dismissed in a 15-minute primary care appointment as “it’s anxiety,” “have you tried meditation,” “let’s talk about your sleep,” or “have you tried losing weight.” They rarely get the response, “Let me refer you to someone who can assess this.”

This is the gap the model is finding. The model doesn’t know you’re a woman in your forties whose mother told her that everyone feels this way. The model doesn’t know that you are good at your job and so you have learned to compensate and mask. The model doesn’t have the bias of a 15-minute appointment in which you only mention the most acceptable version of what’s going on. The model is looking at a longer record than any single appointment can hold, and it is matching patterns the clinician’s schedule didn’t leave time to see.

What I want you to do with this — if you’ve been wondering for a while whether the way you’ve been struggling has a name — is not to chase the model. The model lives in a research paper, not at your doctor’s office. You can’t ask for the score, unfortunately.

What you can do is take the finding as permission. Permission to ask the question out loud. Permission to call a clinician who does adult ADHD assessments — not your primary, the kind of clinician this is the actual job of — and to describe what your life has actually been like, not the version of it you’ve been able to function inside.

I’m not making a diagnosis from a newsletter. The thing I’m pointing at is that the system has been missing this in women like me and you for decades, yet apparently even a relatively simple pattern-matching tool can see it. If a model trained on routine records can flag the pattern, your own description of your life — given to the right clinician — is a stronger signal than anything the model has.

The permission is no longer just from a stranger on the internet. It’s from a research model that sees the pattern faster than the average doctor’s appointment is built to detect it.

A question for you:

When did you first wonder if the way you’ve been struggling has a name?

— Rachel

P.S. — PsyPost has the lay summary; the paper itself is in European Psychiatry. Search the journal for “predictive model adult ADHD electronic health records 2026” if you want the primary source.

Share

THE VIBE

From Vibe Coding to Agentic Engineering

Andrej Karpathy coined the term “vibe coding” in 2025. He ‘retired’ the name in early 2026 and replaced it with “agentic engineering.” Same general idea — humans working with AI agents to build software — with a more deliberate posture. Specs before code. Bounded scope per change. Review every output the agent produces.

The reason for the rename is a number. Per security research summarized in TechTimes’ May 24 coverage, 91.5% of vibe-coded applications shipped in early 2026 with at least one AI-introduced flaw. The same coverage cited a February 2026 industry tracker finding that 63% of vibe-coding users are non-developers.

The wave of new builders is mostly people who didn’t come up through traditional engineering — which is the whole point. And the wave is also shipping software that is structurally insecure most of the time. Both things are true at once. A movement that ships nine of ten apps with AI-introduced flaws cannot survive a second news cycle on its current name. So the name changed, I guess?

Subscribe now

I think the bigger change is the posture under the name.

Vibe coding was permission: permission for a non-engineer to build software without first earning the credentials. Permission to skip the parts of engineering that felt gatekeeping. Permission to ship a working app on a Friday evening because the agent was fast and the feedback loop was solid. The permission was real. A lot of useful software got built that would not have existed otherwise.

Hell, my ADHD app (Codie) and my legal tech startup (Batesly) were vibe coded.1

But the posture also had a cost. The cost was the discipline that traditional engineering builds into a developer’s instinct over years — the reflex to ask “what could break,” the reflex to read the diff, the reflex to write the test, the reflex to imagine the edge case. The agent does not have those reflexes. The non-engineer building with the agent does not yet have them either. So the work shipped without them.

Agentic engineering names the discipline back into the practice. Three things change in how the work gets done:

First, specs before code. Before asking the agent to write the thing, write down what the thing is supposed to do — input, output, edge cases, failure modes. The spec is short. It can be five sentences. But it exists before the agent starts. The agent’s job becomes “implement this spec,” not “figure out what I meant.” That single change cuts the surface area of the bug class the agent introduces.

Second, narrow diffs. Don’t ask the agent to refactor a whole module. Ask the agent to change one function. Read the diff. Read the test. Approve or reject. Then the next change. This feels slow at first, but it really isn’t. It’s the pace at which work that doesn’t have to be redone later actually moves.

Third, review every output. Don’t skim, okay? Read every line. The agent is producing source code that will run on real data and affect real users. If you wouldn’t commit code from a junior engineer without reading it, don’t commit code from an agent without reading it. The agent is a tireless junior engineer who will produce as much code as you let it. The bottleneck is your reading. (See my piece about AI writing code like a confident intern here).

There’s a generational story under this. The non-engineer builder wave isn’t over — I’m sure as shit not going anywhere. It’s shifting from “fast” to “careful and also still fast.” The builders who survive the shift are the ones who absorbed the discipline that the first wave skipped. The ones who don’t will keep shipping the kind of software that the 91.5% number measures.

THE VERDICT

Quick Takes

• The 91.5% number is the reason “vibe coding” had to be retired as a label. A movement that ships nine of ten apps with AI-introduced flaws does not get a second news cycle on a name that signals “we don’t review the code.” The rename is a survival move for the practice. Don’t get me wrong, I’m still using the term — I couldn’t care less if the Tech Bros judge me.

• “Agentic engineering” is the version of vibe coding that survives a security audit. Same same, plus the missing discipline. If you build software and can’t defend the shipped code line by line, you aren’t doing agentic engineering yet. That’s a learnable shift, but it isn’t free.

• The non-developer builder wave isn’t over — it’s sorting. The first wave proved the ‘unlock’ was real. The next wave will sort into the ones who internalize the spec-first discipline and the ones who keep shipping the apps the 91.5% stat is about. The sorting matters more for the careers it builds than for the apps.

  

THE CONFESSIONAL

I’m Still Building the Discipline

The discipline is harder to build than it looks from outside it.

The instinct that wants to skip the spec — the one that says “I know what I want, I’ll just describe it to the agent and edit later” — is a strong instinct. It’s also the exact instinct that produces the 91.5% bug count. Specs feel like overhead though. Until the third time you’ve rewritten the same function because the agent kept solving a different problem than the one you meant, and you finally write the spec, and the agent gets it right the first time, and you realize the spec was the work the whole time. Ughhhh.

Skimming a 200-line diff is not reading it, even if it feels like it. Reading it means reading every line and asking what each line does. That takes longer than skimming. It takes long enough that, when you start doing it, your per-change pace goes down while your per-shipped-feature pace stays the same, because the ratio of reworked changes drops.

I’m inside this shift. I’ve caught myself skipping the spec. I’ve caught myself skimming the diff. I haven’t perfected the posture. The reframe Karpathy named is a real thing to do, and you do it every working day, but you can lose the discipline in a week of speed mode if you let yourself… because ADHD…

If you’re building with an agent and you haven’t felt the discomfort of slowing down on purpose — your turn is coming.

— Rachel

Share

Hey y’all,

Two independent surveys, different methodologies, different geography, same population finding.

• EY’s Global Neuroinclusion at Work Study 2025

  • Surveyed over 2,000 professionals across 22 countries and 8 sectors. It reports that 79% of neurodivergent professionals already use AI at work and are 55% more likely to use AI than their neurotypical colleagues.

• The Understood 2026 Neurodiversity at Work Survey

  • Run by The Harris Poll across 2,073 U.S. adults, found 78% of neurodivergent employees use AI tools versus 59% of neurotypical employees.

If you’re an ADHD adult who started using AI in the last two years and felt something land — something specific, something that wasn’t there before, something you didn’t have a word for — those numbers are naming a population. The thing you felt is felt by most people whose brains work like yours and mine.

Subscribe now

Here’s what I think is happening underneath the numbers.

1. AI is async.

The EY survey reports 71% of neurodivergent users find AI very or extremely effective at facilitating asynchronous working practices. You write when you’re ready. The meeting doesn’t have to have started. The room doesn’t have to be ready. Nobody’s emotional state has to be navigated while you think out loud. Async is the format the ADHD brain has always wanted.

2. The social cost of asking drops to zero.

You can ask the dumb question, ask the same question four ways, change your mind, come back tomorrow and ask the same thing again because you forgot. No friend with a face is wondering why you keep asking. Lowering the social cost of asking is lowering one of the most exhausting taxes ADHD adults pay all day.

3. Revision becomes free.

EY’s earlier collaboration with Microsoft found 80% of neurodivergent users report AI improves their written communication. You can rewrite the draft fourteen times; the fourteenth time, the AI is just as engaged as the first. If you ask a human to read multiple drafts, by draft three, they’re gently and reasonably done with you. Infinite patience matters because perfectionism without infinite patience is a stalemate, and ADHD adults often live inside that stalemate.

4. The system holds the context for you.

The same EY/Microsoft work found that 59% of neurodivergent users report AI improves memory and recall. You don’t have to remember what you said yesterday or re-establish the thread. You can come back to the project after a week of focus drift, and the model will still have everything that was true a week ago. AI lowers the mental-load tax of “where was I” is the tax that ADHD makes the heaviest.

5. And it is a tireless first-draft partner.

AI’s advantage is stamina — it will write a hundred drafts to your one. The bottleneck for a lot of ADHD adults is getting started; the blank page is the enemy. A first draft generated in thirty seconds, that you can edit into something good, functions as an accommodation. The accommodation isn’t being marketed that way, but it’s how it works.

I want to be careful not to over-claim. AI does not fix executive function. It does not fix the cost of sustained focus or the social masking that exhausts neurodivergent adults by 2pm.

But the gap between “I could not do this work” and “I could do this work” is meaningful. And that gap is being closed faster for adults with ADHD than for the general workforce, because the design of the tool happens to fit the shape of the brain.

A question for you:

What’s the first thing AI gave you back?

— Rachel

P.S. — Both reports are public. Search “EY Global Neuroinclusion at Work Study 2025” on ey.com and “Understood 2026 Neurodiversity at Work Survey” on understood.org/work. The agreement between two independent surveys is the part of this that surprised me most.

Share

Hi,

It’s Wednesday. I should be writing to you with something tidy — a small observation from a week that mostly behaved.

I’m writing instead from the residue of Memorial Day weekend, in the particular exhaustion that a lot of women understand and a lot of men do not.

Subscribe now

The official version: three days off. No meetings. No deadlines. Kids around. Open time, on paper.

The actual version: three days of being the operating system for everyone in the house. Breakfast decisions. Sunscreen reapplication windows. Who is bored, who is hungry, who is hot, who is cold, who needs the bathroom right now, versus saying so because the other one said so. Which kid is moving from “playing nicely” to “about to make the other one cry,” and the half-second of intervention you have to land before that crosses over. Snacks. Snacks again. A grocery order because we ran out of the only bread one kid will eat. A round of “who’s going where this afternoon.”

Continuous decision-making with no off switch. O-VER-STIM-U-LATED.

Here’s the thing about ADHD and autism. Executive function — the part of you that gets started, sequences the steps, keeps the right priority on top — runs on external structure. Calendar appointments help. Deadlines help. Other people’s schedules act as scaffolding I clip my own attention onto.

But for me, three-day weekends remove the scaffolding. The kids are home; the schedule is “the schedule we make up as we go”; every decision is fresh; every transition is a negotiation. By Monday night, I was more depleted than I am after a normal workweek, and I had less to show for it.

I used to think the problem was a personality flaw — the kind of person who can’t relax, the kind of person who needs to be told what to do. The actual mechanism is more boring than that. My brain runs on external input. Take the inputs away, and it has to generate them itself, and that generation cost is enormous. At the end of three days of self-generated structure, I am not rested. I am fucking drained.

And I feel guilty for being a mom who was on edge instead of “relaxing” — but it’s really hard being a neurodivergent parent to three neurodivergent kids, okay? I’ve never lived this life before — there’s no instruction manual (other than my decision to do the opposite of my own parents), and every year is my first year having this exact combination of kids at their ages and stages.

Share

On Sunday, a friend sent me this Thread:

We laughed and agreed that these long holiday weekends actually feel like punishment for parents of toddlers (she and I each have a 2-year-old; I also have a 4-year-old and almost-7-year-old). It’s hazing, really. A humiliation ritual.

“On duty” is the closest word I have for what most long weekends feel like once kids and a household and a startup are in the mix at once.

I don’t have a clean ending for this post. I think the recognition is the thing. If you came out of the long weekend feeling like you should be more rested than you are, and instead you came out drained — you were probably on duty. The math is mathing.

A question for you:

What’s the work in your life that gets called “rest” by everyone else and feels like work to you?

— Rachel

Share Vibes & Verdicts

I signed up to be the snack parent for my daughter’s softball team because our local Little League requires parents to volunteer or pay a $50 fee. That is the entire decision tree. I was registering her for softball in the parent portal, and it said, “volunteer for one of these open spots or pay the $50 shame fine.” Okay the “shame” was implied because the whole premise is punishment. But also my ADHD loves to say “yes”.

Subscribe now

There’s no way I was going to volunteer to be a coach, and the only other options were (a) Little League board member, or (b) snack parent. I chose snack parent because organizing a snack rotation among the families is something I can bust out in 10 minutes.

This is the part of the email where I’m supposed to tell you that being snack parent has been a beautiful exercise in community and presence. It hasn’t. I had the parents select which games they’d be available to bring the snacks (via Google Forms) and then ran the answers through Claude to populate a schedule. Easy. But then some parents had qualms with the schedule and others were surprised they weren’t assigned snack duty even though they didn’t fill out the form. Cool.

I’m also, in the same season, trying to ship a piece of software into a regulated industry. The two jobs do not feel like they belong to the same person. One job rewards precision. The other one rewards picking up your phone fast enough to confirm that yes, you are bringing the orange slices. There is no version of me that is good at both at the same time.

The job you said yes to in a moment of impulsivity is still a real job. My ADHD brain wants to treat the softball texts like background noise because they aren’t the “real” work. But they are real. The parents who reply within ten minutes are paying attention. The ones who don’t are being judged, even if no one is saying it out loud. I have been on both sides of that. And it’s weird social dynamics that I don’t respect, to be honest.

The part I didn’t expect: the snack parent thing was the only thing in my life where the deliverable is small and finite.

Confirm the snack. Print the roster. Send the reminder. There is no audit. There is no investor. There are just orange slices, and then it’s done, and the dopamine hit is real.

I think I needed something where the success criteria fit on one line. The startup doesn’t have that. The newsletter doesn’t have that. My law practice definitely doesn’t have that. But the snack rotation does, and apparently, I will take my wins where I can find them.

The softball season ended early for my kid, who decided it wasn’t her thing after a month or so. For some reason, the parents acted like the snack schedule would need to be “redone” just because I wasn’t involved anymore… I still don’t understand the logic there (you all know your assigned dates for the remainder of the season, why are you panicking?)

A question for you:

What’s the small, finite job you took on by accident that turned out to be the only thing in your week with a clean ending? I want to hear about it. Reply to this email — I read everything.

— Rachel

P.S. — We’re back to gymnastics, and the girls have missed the last two weeks because of childcare issues while I’m at work events. I’m telling you this because I’ve been thinking about it and I need to externalize my guilt.

Share Vibes & Verdicts

Two Wednesdays ago, I wrote an essay called “Why I’m Letting You See More of Me” — the first one in this column — and a little while later, I got the first real subscriber reply I’ve ever gotten. Someone I’d never met, cheering me on. I read it three times standing at the kitchen counter.

And then I went silent.

Subscribe now

I haven’t sent you a Margins in two weeks. I haven’t sent you a Workshop since April 19. Same reason for both — I needed a break, and I was using the systems.

I want to tell you what that means. Not because I owe you an explanation — you signed up for a newsletter, not a confessional — but because the reason I went dark is also the reason I’m writing today.

About 6 months ago, I started doing three things at the same time that I had no business doing at the same time. I was preparing to open my freelance law practice. I was building Batesly. I was debugging Codie. And somewhere in the middle of those things, I was also trying to be the kind of mother I want my kids to be proud of — present, organized, not eyeing my phone during school pickup. I was failing at one of those three on most days.

For a while, I told myself the problem was just volume. Three jobs is too many jobs — of course I was failing. But on the afternoons I had a clear two-hour block, I'd still fail at small things. Fifteen minutes to write a LinkedIn post on my way to a networking thing, and the post wouldn't get written. Twenty minutes between client calls to respond to the school email, and the email would sit. So it wasn't the minutes. Something else was wrong.

What I eventually figured out is that every task started from zero — I'd sit down with the fifteen minutes I had and spend the first five rebuilding what I already knew. What my voice sounded like that day, trying to pick up from where I left off. Who I'd already said what to. Which kid had an extracurricular that evening. What my standing answer was to the request for another volunteer gig something-or-other for the PTA. The minutes weren't missing; I just couldn’t get started (it’s hard to be inspired to write when you’re running on fumes).

So I built two AI systems for myself.

One that holds my voice — a Claude project that knows how I write, what I sound like, what I’d never say. Another that holds my life — a Claude project that knows my family, my vendors, my decisions, my rhythms. The first brainstorms with me and the second keeps organized.

The two weeks I spent away from this newsletter, I was using both of them to clear two weeks’ worth of accumulated everything. Email drafts I’d been putting off. School logistics I’d been winging. A memo I’d been promising myself I’d write. LinkedIn posts. An update for Batesly. A vendor renegotiation for the cleaning service. And — I’m noting this without irony — a stretch of actual rest, on purpose, for the first time in a millennia.

The part that I sat on for a while: I think a lot of the women I know would want these too.

I’m offering to build them for you, too.

Two offerings, in plain English

AI Writing System.

I build you a Claude project trained on your voice — your samples, your style, your patterns, your “I would never say that.” It writes LinkedIn posts, About sections, newsletters, recruiter replies, networking notes, conference bios — anything you’d rather not start from scratch every time. It knows you, brainstorms ideas, and drafts content for you to review and edit, saving you time. It gives you a significant jump start without being the automated "AI wrote this generic thing for me.”

• The Build ($2,550, ~one week) is the core buildout.

• The Build + The Reclaim ($3,150) adds a post-event LinkedIn refresh — new headline, About, three posts — done together in the handoff meeting.

• The Upkeep (starting at $950/month) keeps it tuned over time: quarterly voice-sample refresh, new prompts, async support.

AI Chief of Staff.

I build you a Claude project trained on your life — your family, your vendors, your standing decisions, your patterns, your activities. It coordinates the calendar, manages the vendors, drafts the school emails, holds the recurring decisions, runs your standing cadences. The Chief of Staff function your business has, for everything else.

• The Briefing ($1,550) is the strategic diagnostic — pre-call intake questionnaire, a 90-minute consultation, and a written 10–15 page Briefing on what’s working, what’s leaking time, and what a Chief of Staff would change first. Most clients move from The Briefing into The Onboarding. Some take the document and run with it themselves.

• The Onboarding ($7,500, 4–6 weeks) is the full buildout. Briefing memory, calendar coordination, vendor management, decision support, communication drafting, project coordination, standing cadences.

• The Standing ($1,250/month) is the ongoing relationship — monthly working session, quarterly Best Days refresh, new use cases as your life evolves.

Both live on rachelbender.co. Click Work With Me in the nav.

The Writing System is for professionals who write a lot but write the same things over and over. The Chief of Staff is for working mothers (or any parent) running businesses and households.

If you're not sure which one to choose (or whether either is a good fit), there's a free 15-minute discovery call. I'll point you to the right one — or tell you it's not a fit, in which case at least you'll know.

Share

Why this lands in The Margins, not in a Workshop

Because I built both of these for me first. Not as products, but as scaffolding.

The product version exists because enough people have asked some version of how do you actually do all of this? — and the honest answer is, I don’t. The systems do a third of it. That’s a Margins story, not a Workshop story. The Workshop will catch up later with the tactical breakdown of how I built them.

Today is the announcement.

If one of these sounds like you — check out my website link above. If it doesn't, but you know a woman who'd recognize herself in this essay, please forward it to her.

What’s on your mind?

I want to hear from you on this one. If you’ve thought about hiring help — a human assistant, a coach, a chief of staff, a ghostwriter — and bounced off because of cost, fit, or trust, I want to know what stopped you. Reply to this email; I read every one. I’ll likely write the next Margins about what you tell me.

To the subscriber who replied after the first Margins: thank you. Your message was the first real sign that someone was reading. It's part of the reason I came back to this column instead of letting it lapse.

— Rachel

Share Vibes & Verdicts

Why I’m Letting You See More of Me

I almost didn’t send this one.

For two months, Vibes & Verdicts has been a Sunday thing. One email a week. Legal tech, vibe coding, Batesly war stories. Tight scope. Clean category.

And every single week, I’ve sat down to write it and had to cut things out.

Not because they weren’t interesting. Because they weren’t on topic.

The time my kid climbed onto my lap mid-demo and asked “why are the robots sad?” while I was showing an AI-generated contract review to a prospect — cut.

The morning I shipped a feature at 6:14 AM and immediately cried on the kitchen floor because I’d been awake since 3:40 AM and I still had to get the kids ready for school — cut.

The meeting where a male founder, who had never practiced law a day in his life, “explained” contract law to me. I am an attorney. This happened. Cut.

The ADHD-brain moment where I forgot I had a board call and was discovered in a store dressing room, making decisions about a five-figure annual contract on my phone while my kid tried on a Bluey dress. Extremely cut.

I kept cutting because I thought you signed up for legal tech opinions and AI takes, and I didn’t want to waste your inbox on the rest of it.

But the rest of it is the whole thing.

WHY “THE MARGINS”?

In law school, we read everything with a pen in our hand. The good students — the ones who ended up in Biglaw with six-figure clerkships — underlined things in the main body. The great students wrote in the margins.

The margins were where the real analysis happened. The “wait, this doesn’t hold up” moments. The “this conflicts with the case from three weeks ago” connections. The professor’s actual thinking, before it got sanitized into a casebook footnote.

That’s what Wednesday is.

• The Workshop (Sundays) is still the main body. Tactical. On-topic. Built for you to forward to your CTO when you want to argue that vibe coding is real.

• The Margins (Wednesdays) is the pen work. The connective tissue. The part where I tell you what it actually looks like to run a law practice while building a startup while raising kids while having an ADHD brain that treats deadlines like polite suggestions.

You don’t have to read both. You can mute Wednesdays in your settings. I won’t know, I won’t be offended, and the Sunday edition stays exactly the way it was.

But if you want the rest of it — the part I’ve been cutting — it’s here now.

Subscribe now

What it won’t be

It won’t be a mommy blog. I don’t have the patience to write those, and you don’t have the patience to read one from me.

It won’t be polished. I’ll probably send a few I regret. I’ll definitely send one with a typo. I’m writing these on Tuesday nights after the kids go down, and you’ll be able to tell.

But it’ll be honest. That’s the only commitment I’m making.

What I’m actually hoping for

I want the women in this list to stop feeling like they’re the only ones whose “work-life balance” is actually “barely holding it together while pretending to be a professional.” Fuck that.

I want the parents to know that building a company and raising humans at the same time is genuinely insane, and that insanity is a feature, not a bug.

I want the ADHD brains to stop apologizing for the way they work.

I want the people who’ve been told they can’t — by an industry, by a mentor, by a board member, by their own internal monologue — to see what “doing it anyway” actually looks like from the inside.

Mostly I want to stop cutting things.

What’s on your mind?

Hit reply. Tell me something. This column works better if it’s a two-way street, so here’s the first question:

What’s the thing you’ve been cutting out of your own professional brand?

Not the version you put on LinkedIn. The version you’d tell a friend over drinks but would never put in an email blast. The thing that doesn’t fit.

I read every reply. I’ll write about some of them — anonymously, and only with permission.

— Rachel

THE VIBE

AI Writes Code Like a Confident Intern

I need to tell you about the scariest thing I’ve found in my own codebase.

A few weeks into building Batesly, I set up a code review workflow using Claude Code. I wrote a prompt that tells it to scan my entire repository, classify every issue by severity, and fix the critical stuff automatically. I expected it to find some lint errors. Maybe some unused variables. General housekeeping.

It found 47 issues.

One of them was a multi-tenancy data isolation bug. The permission check was filtering by user but not by organization. The app “worked” perfectly in testing — because I was only testing with one organization.

Think about it… the AI wrote code that handled authentication correctly but failed at authorization. It checked “is this a valid user?” but not “does this user’s organization own this document?” The app looked secure. It felt secure. It was not secure.

This is what I mean when I say AI writes code like a confident intern. The output is syntactically correct, follows patterns, handles the obvious cases, and looks professional. But it makes assumptions. It takes shortcuts. It misses the edge cases that only matter when real data and real people are involved.

Here’s a taxonomy of AI code bugs I’ve encountered, ranked by how much they kept me up at night:

• Authorization gaps (terrifying): Missing org_id checks on database queries. Permission checks that verify roles but not resource ownership. API endpoints that authenticate users but don’t authorize the specific action. These look correct at a glance and fail catastrophically in production.

• Silent failure modes (sneaky): Error handling that catches exceptions and does nothing. API calls that fail and return empty arrays instead of errors. Database queries that return no results because of a bad filter, and the UI just shows “no documents found” instead of flagging the problem. The app never crashes. It just quietly gives you wrong data.

• Hardcoded assumptions (annoying): Values that should be environment variables baked into the code. URLs pointing to localhost that work in development and break in production. Default admin credentials that were supposed to be temporary. These aren’t security vulnerabilities per se, but they’re the kind of thing that makes your ops life miserable.

• Optimistic data handling (subtle): Code that assumes data will always be in the expected format. No null checks. No type validation on user inputs. No handling for the case where a related record has been deleted. These bugs hide until someone does something unexpected — which, with real users, happens immediately.

The review workflow I built to catch these is simple: I build features in Replit, sync to GitHub, then run Claude Code against the repository with a comprehensive review prompt. Claude Code identifies issues, classifies them by severity, and fixes the critical ones. I review the fixes, push the clean code back to GitHub, and pull it into Replit.

It’s effective (but not very elegant). And it’s the difference between shipping software and shipping liability.

If you’re vibe coding anything that handles real data — user information, financial records, sensitive documents — you need a review workflow. The AI that writes your code is talented and careless. Just like an intern. And just like an intern, it needs a senior review before anything goes to production.

Subscribe now

THE VERDICT

Quick Takes

• “No-code” platforms are creating a wave of insecure applications and nobody’s talking about it. When you make it easy for non-engineers to build software, you also make it easy for non-engineers to build insecure software. The tools have a responsibility to bake security in by default. Most don’t.

• The duty of technology competence applies to AI-generated code too. Lawyers have an ethical obligation under ABA Model Rule 1.1 to understand the technology they use. If you’re using AI to build tools that handle client data, “I didn’t write the code” isn’t a defense. You chose the tool. You shipped the product. The responsibility is yours.

• Security audits aren’t just for enterprise companies. I ran a comprehensive security audit on Batesly when it was still just me using it. Found and fixed 47 issues. The cost: zero dollars (Claude Code). The time: one evening. The peace of mind: considerable. If you’re building anything that will ever touch user data, do this now, not later.

THE CONFESSIONAL

The Multi-Tenancy Bug Was Live for Weeks

That data isolation bug I mentioned? It wasn’t in some unreleased feature. It was in the production codebase for multiple weeks before I caught it.

No one exploited it. There were no other tenants in the system yet. The exposure was theoretical. But “no one exploited it” is not the same as “it was fine.” If I’d onboarded two firms before running that audit, one firm’s attorneys could potentially have accessed another firm’s documents.

For a legal tech platform. Where the entire value proposition is trust and confidentiality.

I’m sharing this because I think the vibe coding community needs more honesty about the risks. Building fast is great. Shipping features is great. But if you’re not auditing what you ship, you’re gambling. And when the stakes are client data, the house edge is not in your favor.

THE VIBE

I Deployed to AWS Without Knowing What a Docket Container Is

Let me set the scene: it’s a Tuesday night, I’ve been staring at the AWS console for three hours, and I’m reading a documentation page that casually uses the phrase “orchestrate your containerized workloads across a fleet of EC2 instances.” I understand every word individually. Together, they mean nothing to me.

This is the part of vibe coding nobody talks about. The building part is fun. The deploying part is where you question every decision that led you here.

Here’s the thing: Replit is great for development. You write prompts, features appear, you test them in a live preview. It feels like magic. But at some point, if you’re building something real — something that handles client data, needs to be secure, needs to scale, needs to not go down when Replit has an outage — you need actual infrastructure.

For Batesly, that meant AWS. Specifically: RDS PostgreSQL for the database, S3 for document storage, and App Runner for the application itself. Here’s how I got there and what I wish I’d known.

The database was the easy part. RDS PostgreSQL is basically “managed database as a service.” You click some buttons, pick a size, set a password, and you have a production database. Connecting it to my Replit app meant putting the credentials in environment variables and swapping the connection string. Total time: about two hours, most of which was me being nervous about clicking the wrong thing.

S3 was surprisingly intuitive. Create a bucket, name it, enable encryption, block public access. Done. The concept is dead simple: it’s a folder in the cloud. Files go in, files come out. The part that took me longer was understanding presigned URLs — temporary, secure links that let users upload and download files without ever having direct access to the bucket. Once I grasped that concept, everything clicked.

App Runner was the wildcard. This is where Docker containers live, and I’m going to be honest: I still only sort of understand what’s happening. Here’s my working mental model — a container is like a box that has your app and everything it needs to run inside it. App Runner takes that box and runs it on AWS’s servers. You don’t need to know what’s inside the box works at a low level. You need to know that the box is configured correctly.

The biggest surprise was how cheap it all is. My entire AWS bill for a production-grade infrastructure — encrypted database, secure file storage, application hosting with auto-scaling — runs about $47/month. The database is the big expense. S3 storage is pennies. App Runner charges for compute time.

For context: that’s less than one seat of most legal tech software. I’m running an entire platform for less than what some firms pay for a single user’s access to their DMS.

What I’d do differently: I’d set up separate environments (dev, staging, production) from day one. I didn’t, and now I’m doing surgery on a live patient every time I push an update. It’s fine when you’re the only user. It’s going to be a problem very soon. Also: I’d learn basic AWS IAM (identity and access management) before touching anything else. Understanding who has permission to do what is the foundation of everything in AWS, and I stumbled through it instead of learning it properly.

The meta-takeaway: deploying to real cloud infrastructure as a non-engineer is doable. Not easy — doable. The documentation is dense, the terminology is hostile, and the console has approximately nine thousand menu items. But the core concepts are learnable, and once your stuff is running, it’s genuinely more reliable than any “easy” deployment platform.

You don’t need to understand everything, you need to understand just enough. And “enough” is a smaller number than the industry wants you to believe.

Subscribe now

THE VERDICT

Quick Takes

• AWS documentation is written for people who already know AWS. This is the single biggest barrier to entry for non-engineers. The docs assume you understand networking, IAM, VPCs, and security groups before you start. A genuinely beginner-friendly AWS onboarding flow would be worth billions. Someone should build it. (Not me. I’m busy.)

• The “just use Vercel/Netlify” crowd isn’t wrong, but they’re not right either. For marketing sites and simple apps, absolutely. For a multi-tenant platform that handles sensitive legal documents with encryption requirements and audit logging? You need more control than a one-click deploy gives you. Know when you’ve outgrown the easy path.

• The real cost of building software is time, not infrastructure. My AWS bill is $47/month. My opportunity cost of the hundreds of hours I’ve spent building instead of billing clients is… a number I don’t like to calculate. The infrastructure is cheap. The founder’s time is expensive. Budget accordingly.

THE CONFESSIONAL

I Broke Production on a Wednesday

There’s a special kind of panic that comes from pushing an update and watching your live application crash. Not a “something looks weird” crash. A “the entire page is a white screen and the console is screaming” crash.

It was a database migration that went sideways. I’d changed a column type without accounting for existing data, and the app couldn’t read the table anymore. The fix took twenty minutes. The cold sweat lasted considerably longer.

This is why staging environments exist. This is why people test migrations before running them on production. I knew this intellectually. Now I know it viscerally.

Nobody was using the app at the time except me, so the damage was limited to my blood pressure. But it was a wake-up call: the gap between “building a product” and “running a product” is wider than I thought. Building is creative. Running is operational. And operational means not breaking things that are already working.

THE VIBE

Your Favorite Legal Tech Company Hates You

I need to talk about something that’s been bothering me since I started researching the legal tech market: the pricing.

Not in a “things are a bit expensive” way. In a “this is genuinely predatory and everyone has just accepted it” way.

Here’s what a typical mid-size law firm’s tech stack looks like in 2026: iManage or NetDocuments for document management ($30–60/user/month). Clio or PracticePanther for case management ($39–89/user/month or more depending on the tier). Separate billing software. Separate trust accounting. Separate eDiscovery tools when litigation heats up. An email integration that sort of works. A client portal that nobody likes (seriously — it’s universal).

For a 25-lawyer firm with support staff, you’re looking at $8,000–$15,000 a month. For tools that don’t talk to each other. That require separate logins. That store your data in separate silos. That you can’t leave because the migration cost is astronomical and they know it.

The per-seat model is the original sin of legal tech. It doesn’t charge you for value, it charges you for headcount. Hire a new associate? That’ll be another $150/month across your stack. Bring on a paralegal? More money. A temp for a big case? Hope you like per-seat overages.

It’s a business model that punishes growth. And the vendors have figured out that the switching costs are so high — migrating a document management system with 100,000 files and years of metadata is a genuinely miserable experience — that they can raise prices and degrade service and you’ll stay anyway.

I don’t say this as a neutral observer. I’m building a competitor, so I’m obviously biased. But I’m also a practicing attorney who has watched firms I’ve worked with hemorrhage money on tools that haven’t meaningfully improved in a decade. The bias comes from the frustration, not the other way around.

Here’s what I think the market is missing:

• One platform, flat fee. Documents, matters, billing, trust accounting, client portal, eDiscovery — all in one system. $299 to $1699 a month depending on firm size. Not per seat. Not per feature. The whole thing.

• Data portability as a feature, not a hostage negotiation. You should be able to export your entire dataset at any time in standard formats. If I can’t keep you by being better, I don’t deserve to keep you at all.

• Architecture that actually integrates. Not “integrations” where you pay extra to connect two products that should have been one product. Built-together, same-database, shares-the-same-permission-model integration. When you upload a document, it’s immediately searchable, automatically tied to a matter, and accessible to everyone who should see it — because it’s all one system.

Is this ambitious? Absolutely. Am I building it alone with AI tools and a budget that wouldn’t cover one enterprise sales rep’s quota? Also yes. But the infrastructure costs for building software in 2026 are a fraction of what they were five years ago, and that changes the math on what’s possible.

The legal tech incumbents aren’t bad companies run by bad people. They’re mature businesses optimizing for revenue extraction from a captured market. That’s rational. But it creates an opening for someone willing to build something better and charge less for it.

I’m not the only one who sees this. I’m just the one dumb enough to try.

Subscribe now

THE VERDICT

Quick Takes

• If your software vendor makes migration painful on purpose, that’s not a feature. I’ve talked to firms who’ve stayed on terrible platforms for years because the migration estimate was $50K+. That’s a hostage situation, not loyalty. Any product that can only keep customers by making it expensive to leave is telling you something about its value.

• The “best of breed” argument is over. For years, the pitch was “use the best tool for each job and integrate them.” In practice, this means six tools, six logins, six billing cycles, and an IT person whose entire job is keeping the integrations from breaking. All-in-one wins when the “one” is good enough. And “good enough” keeps getting better.

• Bookkeepers think in ledgers, not UIs. Something I learned preparing for a trust accounting demo: when I built documentation showing pretty screenshots of the interface, the firm’s bookkeeper didn’t care. She wanted to know how data flows between accounts. She wanted to see the reconciliation math. The lesson: know your audience. Not everyone evaluates software the same way.

THE CONFESSIONAL

I Priced Batesly Wrong Three Times

• First attempt: $99/month flat. Too cheap. People literally told me they wouldn’t trust legal software that costs less than their Netflix subscription. Pricing communicates value, and I was communicating “side project.”

• Second attempt: Per-seat pricing to match the market. This lasted about 48 hours before I realized I’d become the exact thing I was building against. Hypocrisy speed run.

• Third attempt (current): Tiered flat-fee. $299 for small firms, $599 for mid-size, $999 for larger firms, $1699 for enterprise. Unlimited users at every tier. This feels right.. but I do plan to increase each tier after the Big Launch (early adopters get the lower prices locked-in for life). It’s affordable enough to be a no-brainer upgrade, expensive enough to be taken seriously, and the flat-fee model is a genuine differentiator in a market addicted to per-seat.

• Pricing is strategy. It took me three tries to figure out mine. If you’re building something and you’re agonizing over pricing, just pick something, put it out there, and listen to what people tell you. The market will teach you faster than any spreadsheet model.

  

THE VIBE

The Prompt is the Product

Here’s the thing nobody tells you about vibe coding: the quality of what you build is directly proportional to the quality of your prompts. Not your coding skill. Not your technical knowledge. Your prompts.

I learned this the hard way. Early on, I’d write prompts like: “Build me a document upload feature.” And I’d get something that technically uploaded documents. It would also have no error handling, no file size limits, no virus scanning, no audit logging, and a UI that looked like it was designed by someone who’d never used a computer.

The AI did exactly what I asked. The problem was what I asked.

Now my prompts look completely different. Here’s a real one from last week: “Build a secure document upload flow for a legal document management system. Requirements: presigned S3 upload URLs with 15-minute expiry, file size limit of 100MB, allowed MIME types are PDF, DOCX, XLSX, PNG, and JPG. On upload confirmation, record document metadata in the documents table including SHA256 hash, file size, MIME type, and uploader’s user ID. Write an entry to the audit_log table with action type DOCUMENT_UPLOADED. The upload should be scoped to a specific matter and inherit the matter’s permissions. Show a progress indicator during upload and a success toast on completion. Handle errors gracefully with user-friendly messages.”

Same feature. Completely different result. The difference isn’t that I learned to code — it’s that I learned to think like a product manager.

Here’s what I’ve figured out about writing prompts that produce production-quality code:

Be specific about the outcome, not the implementation. I don’t tell the AI which React hooks to use or how to structure the API call. I tell it what the user should experience and what the system should enforce. “The user drags a file onto the upload zone and sees a progress bar. If the file exceeds 100MB, they see an error before the upload starts.” The AI figures out the how. I own the what and the why.

Name your constraints explicitly. Security requirements. Performance expectations. Accessibility needs. Error states. If you don’t mention it, it won’t exist. AI doesn’t infer security requirements. It doesn’t add audit logging unless you ask. It won’t check permissions unless you tell it to. Every constraint you care about needs to be in the prompt.

Include context about the system. My best prompts start with a few sentences about what Batesly is and what the relevant data model looks like. “Batesly is a multi-tenant legal document management platform. Each document belongs to a matter. Each matter belongs to an organization. Users can only access matters they’re assigned to.” This framing prevents 80% of the bugs I used to get.

Describe the unhappy paths. What happens when the upload fails? When the user’s session expires mid-upload? When two people upload a file with the same name to the same matter? If you only describe the sunny-day scenario, you get sunny-day code. Real software lives in the edge cases.

Send one prompt at a time. I learned this through pain. When I’d send Replit Agent a massive prompt with ten features, it would get confused, create circular dependencies, and break things that were already working. One feature per prompt. Confirm it works. Sync to GitHub. Then the next one. Sequential beats batch every single time.

The meta-lesson here is that prompting is a skill, and like any skill, you get better with reps. My prompts from three months ago are embarrassing compared to what I write now. Not because I took a course. Because I shipped a hundred features and noticed what worked.

If you’re vibe coding and your results feel mediocre, it’s probably not the AI. It’s the prompt. And that’s actually good news — because it means the fix is within your control.

Subscribe now

THE VERDICT

Quick Takes

• Replit Agent vs. Claude Code: different jobs. I use Replit Agent to build features and Claude Code to review them. Trying to use one for both is like using a hammer to paint. People keep asking me which is “better.” They’re not comparable. One creates. The other critiques. You need both. Having said that, in the past week, I’ve started having Claude (chat) write the prompts for Claude Code, then push to GitHub… cutting the Replit Agent out of the equation entirely. I’ll report back next week.

• The “prompt engineering” job title is already dead and I think that’s fine. It was never a job — it’s a competency. Like “googling things effectively” was a competency in 2008. The people who prompt well will just be better at their actual jobs. The people who made it their whole identity are going to have a rough year.

• Law firms that ban AI are going to lose associates. I’m hearing from more and more young lawyers who are quietly using AI for research, drafting, and review — and lying about it because their firms have blanket prohibitions. The firms that figure out responsible AI policies first will win the talent war. The ones that stick their heads in the sand will wonder why everyone under 35 left.

  

THE CONFESSIONAL

I Used a Fake Statistic in a Pitch Deck

Not on purpose. But still.

A few weeks ago I was building a sales deck and I asked Claude to help me find statistics about legal tech spending. It gave me a beautifully formatted stat: “67% of mid-size firms report dissatisfaction with their current DMS.” Sounded great. Felt right. I put it in the deck.

Then, right before a demo, something nagged at me. I went to verify the source. Couldn’t find it. Because it didn’t exist. Claude had hallucinated a statistic and I’d almost presented it to a 50-attorney law firm as fact.

Now I verify every single data point before it goes into anything external. Lesson learned the uncomfortable way. AI is a fantastic collaborator and a terrible fact-checker. Don’t outsource your credibility.

THE VIBE

I’m a lawyer who builds software and I have no idea what I’m doing. Sort of.

Six months ago, I was laid off from a tech startup. It was my first job in tech, and my first experience with layoffs. (I’d been in private practice for the first 8+ years of my career).

When it happened, I texted a friend who'd worked at all the Big Tech companies. She said, “Is this your first layoff?” — in other words, “welcome to tech.”

My role was Product Counsel; I was the first person to hold the title. I’d only been at the company for a little under a year and my friend explained that when it comes to layoffs, “last one in, first one out” is standard — i.e., the most recent hires are the first ones canned.

At first I didn’t know what to do with myself, but about 3 weeks into unemployment, I discovered vibe coding. My husband and I were on a double date when my husband’s friend (who is well known in Seattle AI circles) showed me an app he built using vibe coding. I was fascinated.

Over the next 3-4 weeks, I played around with Lovable. Then I built Codie — a productivity app for ADHD brains. I also started building Indigogh, an art marketplace with a built-in philanthropic model. I spent a lot of time working on Indigogh before pausing the project to focus exclusively on Batesly.

It’s been 4 months since then, and 2 months since I first announced my Batesly project on LinkedIn. Today, I have a production application running on AWS, 8 law firms in my demo pipeline, and a mass of infrastructure I built by describing what I wanted to an AI and watching it write the code.

I still can’t write a for loop from memory. I’m not sure I ever will. And I’m starting to think that doesn’t matter.

Subscribe now

It’s called Vibes & Verdicts because that’s basically my entire process. The vibe is the building — prompting AI tools, stitching together platforms, shipping features at 1am because something finally clicked. The verdict is the lawyer brain that won’t stop evaluating everything: Is this actually good? Is this secure? Would I trust this with client data? Does this solve a real problem or am I just building to build?

Both parts are essential. The vibe without the verdict is a prototype that never becomes a product. The verdict without the vibe is a 47-page requirements document that never becomes anything at all.

A little more about me… I’m Rachel. I’m a Washington-licensed attorney. I run Arthouse Legal, a fractional general counsel practice for creative professionals. And I’m building Batesly — a legal practice management platform designed to replace the fragmented, overpriced, soul-crushing tech stack that most law firms are trapped in. iManage for documents. Clio for case management. A different app for billing. Another one for discovery. None of them talk to each other effectively and they all charge per seat.

I’m building one platform that does all of it, at a flat monthly rate, with a matter-native architecture that organizes everything around how lawyers actually work — not how software companies think lawyers should work.

I built it using Replit, Claude (chat and code), and AWS, using my workflow: vibe coding — which is what the cool kids call describing what I want in plain English and AI writes the code. Then I review it, break it, fix it, and ship it. The entire production infrastructure cost me under $5K.

Every week in this newsletter, I’ll share what I’m actually doing. Not the highlight reel. Not the “10x founder productivity hack” stuff. The real thing — the tools that work, the ones that don’t, the bugs that make me question my life choices, and the moments where something clicks and I remember why I started.

If you’re building something — an app, a side project, a business — without a traditional engineering background, this is for you. If you’re a lawyer wondering whether AI is going to eat your career or save it, also for you. If you just like watching someone figure things out in public, welcome. I hope you brought popcorn

Let’s do this!

THE VERDICT

Quick Takes

• Vibe coding is getting mainstream attention, and most of the discourse is terrible. Hot take: the people saying “vibe coding isn’t real engineering” are correct. It’s not. It’s something else entirely, and that’s the point. Not everyone building software needs to be an engineer. Some of us need to be domain experts who can direct AI to build what we know should exist. Stop being a weirdo gatekeeper.

• Legal tech pricing is a scam and everyone knows it. Per-seat pricing for software in 2026 is indefensible. It punishes firms for hiring. It makes growth more expensive. And the vendors know you’re locked in because migrating a DMS is roughly as fun as a root canal. Flat-fee pricing isn’t a gimmick — it’s just how software should work when you actually respect your customers.

• The best AI tool is the one you actually use consistently. I’ve tried Claude, GPT, Gemini, Cursor, Copilot, and about fifteen other things. My actual stack is embarrassingly simple: Claude for everything strategic and complex, Replit Agent for building features, Claude Code for reviewing them. Stop tool-shopping. Claude is the best. (No, Anthropic isn’t paying me to say that, but they should.)

THE CONFESSIONAL

‘Being vulnerable is brave’, I say to myself

I almost didn’t start this newsletter.

Not because I don’t have things to say — clearly I do — but because there’s a voice in my head that says: “Who are you to write about building software? You’re a lawyer. You don’t even know what a Docker container is.”

And that voice is technically correct. I deployed my app to AWS and I’m still not 100% sure what a Docker container does versus what App Runner does versus what an EC2 instance is. I know my app runs. I know it’s secure. I know the data is encrypted. I’m a little fuzzy on the exact container orchestration situation.

But here’s what I’ve realized: that voice — the one that says you need to understand everything before you’re allowed to do anything — is the reason most good ideas die. Not because they’re bad ideas. Because the person who had them didn’t have the “right” credentials to try. Fuck that.

So here I am, trying anyway, in public. For all of y’all to judge me. With a newsletter memorizing everything.

Gods help us all.