::case
Aria · 38 · former magazine editor turned indie newsletter writer · 11K subscribers · Brooklyn
She quit her staff editor job at a national magazine in 2024 to launch a Substack. By the time she hit 11K subscribers, she was working 60+ hours a week to keep the publication schedule and was drafting on the subway home from her side consulting work.
::before AI
She wrote everything in Google Docs. Research was browser tabs. The Substack mailing-list management lived in her head and a chaotic CRM. Her newsletter calendar had been 'next Tuesday' for three weeks.
::the stack
- ▲Claude · all long-form drafting and editing (Sonnet 4.5+)
- ▲Perplexity · live web research and primary source discovery
- ▲ChatGPT · subject line A/B/C/D variants and short copy
- ▲Whisper · transcribing the conversations she records on her walks
- ▲Substack native AI · post optimization (already paid for, free with subscription)
::the shift
She built a Skill Primer for her voice. It lives in a file. Every Claude session loads it first. The model now finishes her sentences the way she would, not the way the model defaults to.
::what changed
- ·Drafting time from 6 hours per long piece → 90 minutes (the writing IS still hers; the friction is gone)
- ·Newsletter cadence from inconsistent → twice-weekly, eight weeks running
- ·Research time from 'a whole Wednesday' → 30 minutes with Perplexity → primary sources verified manually
- ·Open rate from 38% to 52% after she started running 12-subject-line A/B tests
- ·Subscribers: 11K → 19K in 4 months (mostly from finally being publishable on her promised cadence)
::the trap they hit
The first month, she let Claude draft a piece in her voice and just lightly edited. The voice was 'close enough.' Three readers emailed: 'something feels off this week.' Trust signal slipped.
::the fix
She switched to: outline with Claude, write the body herself, then Claude does the final critique pass as 'hostile senior editor.' The voice is 100% hers again and the drafting still took half the time.
::one year later
The newsletter is her primary income. She hired one part-time editor who knows the AI stack. She's working 35-40 hours a week now and shipping at 2× her staff-job pace.
“AI didn't write my newsletter. It removed the friction between my brain and the page. There's a difference.”
::case
Marcus · 29 · indie developer · solo SaaS · ~$8K MRR · São Paulo
He built a small B2B SaaS tool serving European retailers. Solo. Bootstrapped. Customer service is him. Marketing is him. Code is him. The product is profitable but he's the bottleneck on every dimension.
::before AI
Customer support emails ate his mornings. New-feature roadmap was a Notion doc he hadn't opened in 6 weeks. He was burning out on his own success.
::the stack
- ▲Claude · code review, architecture decisions, customer support email drafts
- ▲Cursor · in-editor AI for the actual code-writing
- ▲ChatGPT · second opinions when Claude rate-limits
- ▲Local Ollama · for any customer data he didn't want hitting cloud
- ▲Linear AI · sprint planning and ticket triage
::the shift
He wrote AECode contracts around customer-support email patterns. Now Claude drafts each support email in his voice with the right tone, references the customer's history, and proposes the next step. Marcus reviews, edits 2-3 lines, sends.
::what changed
- ·Support email response time from ~24h to under 2h
- ·Customer satisfaction (real CSAT survey) from 78% to 91%
- ·Code review cycles from 'I'll get to it' → same-day, every PR
- ·He shipped 3 features in February that would have been Q2 work the old way
- ·Time he actually has after work: from zero to 'evenings exist now'
::the trap they hit
Month two, he let Cursor auto-apply diffs without reading them carefully. A subtle data-loss bug shipped. Two customers lost recent records. He had to apologize personally and refund.
::the fix
Now every Cursor diff gets a 'hostile principal engineer' Claude review before he approves it. He treats AI-generated code like code from a smart but tired junior dev. Reads everything.
::one year later
He's at $22K MRR. He still hasn't hired. He doesn't want to. He's running an indie SaaS at the scale that used to require a 4-person team. He's traveling more.
“The bug taught me the actual rule: AI is the typist, I'm the engineer. The bug is mine. The receipt is the proof.”
::case
Ms. Reyes · 44 · 7th-grade English teacher · public school · Detroit · 17 years experience
She teaches 5 sections of 7th-grade English. 130 students total. The district adopted a new curriculum in fall 2025 that doubled the writing-feedback load. She was working until 11 PM most weeknights grading rough drafts.
::before AI
Lesson planning happened Sundays. Feedback on student drafts was a slog of marginalia. Parent communication slipped because of the feedback workload. Her family noticed.
::the stack
- ▲Claude · lesson planning and differentiated worksheet generation (no student PII)
- ▲ChatGPT · variant generation for parent communication
- ▲Khan Academy AI · for math-adjacent vocabulary lessons
- ▲The district's vetted AI grading-assist tool · for actual student work (FERPA-compliant)
- ▲Nothing local — she's not technical enough to set up Ollama and that's fine
::the shift
She built a Skill Primer for each unit she teaches. Voice, learning objectives, common student misconceptions. New Claude sessions load the right primer for the right week.
::what changed
- ·Lesson planning time from 6 Sundays/year (the bad ones) to 90 minutes weekly
- ·Differentiated worksheets — she now ships 3 levels of every assignment instead of 1
- ·Parent communication response time from 3+ days to same day for non-urgent
- ·She was back at the dinner table by 6 PM most nights by November
- ·Standardized test scores in her sections rose modestly — likely the differentiation, possibly the increase in actual feedback quality
::the trap they hit
October, she pasted a paragraph of a struggling student's writing into Claude for a quick feedback draft — including the student's name. She caught it before sending the AI response. But the data was already in Claude's chat history.
::the fix
Now she has a strict rule: every prompt about a student gets de-identified first. 'The student' or 'this 7th grader.' She uses the district's FERPA-compliant tool for anything that needs to include identifying information. Took a week to make it muscle memory.
::one year later
She's still teaching. She's not burned out anymore. She gave a presentation to the district on the FERPA-safe AI workflow she uses. They're piloting it across 4 schools.
“AI didn't replace the teacher work. It replaced the part of teaching that was eating my evenings.”
::case
David · 51 · solo practitioner · real-estate + estate planning · midwest small city · 23 years
He has a steady book of business — 60% real estate closings, 40% wills and trusts. Mostly local. He'd watched the AI tools roll into his bar association's CLE materials with suspicion. His paralegal retired in March 2025 and he didn't want to hire.
::before AI
He did everything. Title work, document drafting, client meetings, billing. He was good but slow. Margins were okay; volume was capped by hours-in-the-day.
::the stack
- ▲Westlaw + Lexis · the actual research (verified, billable, malpractice-defensible)
- ▲Claude · document review, plain-language client communication drafts, issue-spotting
- ▲Local Ollama · for any work involving privileged client matter (Llama 3.1 70B)
- ▲His firm's case management AI · for billing review
- ▲NOT ChatGPT — he doesn't trust the data residency
::the shift
He uses Claude as a research associate: issue-spotting before drafting, document review checklist, plain-language client email drafts. He NEVER cites a case Claude gives him without pulling it in Westlaw.
::what changed
- ·Document drafting time on standard estate plans from 4 hours to 90 minutes
- ·Issue-spotting on complex real estate closings is more thorough — Claude catches things he'd miss when tired
- ·Client communication response time from 'within a few days' to same day
- ·He took on 20% more matters in 2025 with no support staff
- ·Bills out at higher hourly because the deliverable is sharper and faster
::the trap they hit
He cited a state appellate case in a memo. Claude had fabricated it confidently — the name was plausible, the citation format was correct, the case did not exist. He caught it in Westlaw before filing. But he came within 30 minutes of putting his bar license on the line.
::the fix
Now every citation gets pulled. Every quote gets pulled. No exception. He has a stamp on his desk that reads 'PULL THE CASE.' His malpractice carrier knows about his AI workflow now and is fine with it because of the verification discipline.
::one year later
He's still solo. He's profitable in a way he hadn't been in 10 years. He gave a CLE talk on 'Solo Practitioner AI Discipline.' The bar association published it.
“The malpractice carrier doesn't care that the AI hallucinated. They care that I cited it. I never will.”
::case
Jean · 71 · retired elementary school principal · widowed in 2023 · suburban Texas · writes a memoir nobody asked for
She retired in 2021. Her husband died in 2023. She started writing a memoir partly to grieve, partly to leave something for her grandkids. She'd never used AI. Her grandson set up ChatGPT on her tablet during a Thanksgiving visit in 2024.
::before AI
Memoir was 80 pages of handwritten notes. She'd type them up in fits and starts. She didn't trust her writing. She didn't trust her memory of dates. She'd given up twice.
::the stack
- ▲Claude (free tier on her tablet) · memoir editor and research aide
- ▲Perplexity (free) · when she needs to verify a date or place from her memory
- ▲She does not have any other AI. She does not need any other AI.
::the shift
She types up a memory in her own voice. Pastes it to Claude. Asks: 'Help me find one detail I might be misremembering. What's a question that would help me deepen this scene?' She doesn't let Claude write. She lets it ask her better questions.
::what changed
- ·Memoir is at 240 pages
- ·She's writing 4-6 days a week, mostly mornings with coffee
- ·She has a sustained creative project for the first time in 30 years
- ·Her relationship with her grandson — who set this up — has a new texture (they email about her chapters)
- ·She found out her grandfather served on a ship that's actually documented in maritime records, which she'd doubted for decades
::the trap they hit
In month two she let Claude rewrite a paragraph 'in her voice.' It came back cleaner, smoother, and not at all hers. She felt a flicker of doubt about her own writing for a week.
::the fix
Her rule now: 'You may critique. You may ask. You may not write. I'm the writer.' Stuck a sticky note on the tablet that says exactly that. The memoir is hers again.
::one year later
She's a quarter through final revision. Her grandson has set up a print-on-demand workflow for when she's done. She's already planning what to write next.
“I'm not afraid of AI. I'm afraid of writing the wrong memoir because the AI was too tempting to say no to. The rule on the sticky note keeps me honest.”
::case
Daniela · 32 · ICU nurse · 8 years experience · large urban hospital · charge nurse two shifts a week
Twelve-hour shifts. Documentation is the silent killer of bedside time. Patient education materials are dated and generic. Her hospital adopted a vetted internal LLM (Azure OpenAI with BAA) in late 2024 specifically for clinical work.
::before AI
She'd stay 45-60 minutes past her shift charting on most days. Patient education conversations were rushed because there was no time to prep the right materials for each patient.
::the stack
- ▲Hospital's internal LLM (BAA-covered Azure OpenAI) · for any PHI-adjacent documentation
- ▲Claude (personal, free tier) · for generic patient-education materials in plain language (no PHI ever)
- ▲UpToDate + Lexicomp · ALWAYS the source of truth for clinical content
- ▲Perplexity · for staying current on guideline changes (verified against official source)
- ▲Whisper (in the hospital's vetted app) · for voice-to-text on shift handoffs
::the shift
She prepares patient education materials on her commute home or at home, in advance, for the conditions she sees most often. Then she has 5 versions ready to print or text to families when she meets a new patient.
::what changed
- ·Off-the-clock charting reduced — she now leaves on time most shifts
- ·Patient education quality went up — families can actually understand the discharge instructions
- ·She started a peer-shared library of patient-ed materials with two other ICU nurses
- ·Charge-nurse handoffs are sharper — she has a structured SBAR template the AI helps her tighten in real-time (no PHI in cloud — the LLM is the hospital's)
- ·Her own resilience: leaving on time, having dinner with family — non-trivial
::the trap they hit
Early on, she pasted a question about a specific patient's lab values into personal ChatGPT (off-shift, on her phone). She caught it before hitting send because the lab values were unusual enough to be identifying. She felt sick about it for a week.
::the fix
Hospital's internal LLM is now the ONLY tool that touches anything patient-identifying. Personal AI is for generic content only. She also taught two newer nurses the rule the same way she learned it: 'Pretend a journalist will read this prompt tomorrow.'
::one year later
She's still in the ICU. She's a quietly known go-to for the hospital's AI-in-nursing rollout. She's not burned out. The work hasn't gotten easier; the documentation overhead has.
“AI didn't save lives in my unit. It gave me back 45 minutes a shift to actually be at the bedside. That's where the saving-lives happens.”
::case
Hector & Rosa · husband and wife · own a 14-table Mexican restaurant in Mesa, AZ · second location opening fall 2026
They opened the first location in 2018. By 2024 they were profitable but barely scaling — every operational decision had to flow through them. Marketing was Rosa's mom posting on Facebook. Bookkeeping was Hector at the kitchen counter at midnight.
::before AI
They thought AI was for tech companies. Their accountant — who is sharp — pushed them in March 2025: 'You're paying me hourly for stuff Claude could draft. Let me show you.'
::the stack
- ▲Claude (free tier · upgraded to paid after month two) · drafts for everything: vendor emails, hiring posts, social posts, menu descriptions, customer review responses
- ▲ChatGPT · second opinion when Claude rate-limits
- ▲Perplexity · checking local food cost trends and competitor pricing
- ▲Imagen · for occasional menu illustration concepts (they still hired a local artist for the final menu art)
- ▲Their POS's native AI · for some inventory forecasting (already paid for in the subscription)
::the shift
Rosa's mom got promoted from 'doing the Facebook' to actually managing the social cadence. Rosa drafts the week's content in 30 minutes on Tuesday morning using Claude. Three posts per channel. Scheduled. Done.
::what changed
- ·Social engagement up 4× — partly cadence, partly quality (the posts are actually about the food now, not generic 'visit us today')
- ·Hiring time-to-fill from 6 weeks to 11 days (Claude writes the listings, screens replies, generates the interview questions)
- ·Bookkeeping prep time on Hector's part down 60% — he drafts the categorization with Claude, accountant verifies
- ·They put together a vendor RFP for the second location in 2 days instead of 2 weeks
- ·They opened the second location on schedule — and on the original budget — partly because operations didn't bottleneck on them
::the trap they hit
Hector pasted last quarter's payroll spreadsheet (with employee names and bank account routing numbers) into ChatGPT to ask 'where can I optimize labor costs.' He didn't realize the data was in the cloud. The accountant flagged it during their April meeting.
::the fix
They now have a written rule taped to the office wall: 'No employee, customer, or banking data into any AI. Ever.' For payroll questions, they use ChatGPT with the data redacted to roles and dollar amounts only ('Cook A: $X, Cook B: $Y'). For anything with names, the accountant's tools.
::one year later
Two locations open. They hired a part-time operations manager. Hector is no longer at the counter at midnight; he's home with the kids. Rosa is sleeping at night. The food is still made by them in the kitchen. That part hasn't changed.
“We didn't expand because we got bigger. We expanded because we stopped doing the parts of the business that didn't have to be us.”