My AI Renaissance

Mid-January 2026, I came back to Claude Code. I had been the Chief AI Officer at Stone Co. for almost three years. My team of ~100 people had built AI systems at real scale — 100% of our customer service and sales running on proprietary AI, nearly R$1 billion in results over a five-year horizon. I knew AI. I was confident. And within a single session, I felt like I'd gone from pole position to dead last.

Not because the code I was writing was bad. Because I realized how little of it I actually needed to write.

Over the next six weeks — squeezed between board meetings, team reviews, and the ordinary chaos of leading a large technology organization — I built twelve pieces of software. Some useful. Some silly. Some genuinely impressive, if I'm honest. This is the story of what I built, what broke, what surprised me, and what all of it implies for the future of software development.


A bit of context

I've been writing code for almost 30 years. I started as a kid, became an entrepreneur in my 20s, built teams and companies through my 30s, and never stopped coding. When I joined Stone three years ago to build the AI team, I brought that coding instinct with me — not because a CAIO needs to code, but because you can't lead something you haven't done yourself.

Our AI work at Stone had been focused entirely on building software that uses AI to solve customer problems: conversational agents, LLM-powered decision systems, retrieval pipelines. Not AI on the SDLC itself. That was the CTO's domain. I'd been experimenting with coding assistants since 2022 — Copilot, Cursor, various agentic modes. I'd shipped a few things with them. But the experience was always the same: a lot of back-and-forth, quality that degraded as context grew, and a mental model of the process that hadn't really changed. I was still the one doing the work; the AI was autocompleting it.

Then I sat down properly with Claude Code. And something actually shifted.

It wasn't that the code was perfect. It's that the interaction model was different. I described what I wanted at the product level. It built it. I refined. It adapted. Context held. The back-and-forth that used to happen inside my own head — "should I use this library, how should I structure this, what are the edge cases" — now happened as a conversation. And the conversation was fast.

I started staying up late. Not because I had to. Because I couldn't stop.


What I built

Here's the full list. Twelve projects in six weeks, all built in evenings and on weekends. Most are running in production today — on my home server, on my Mac, in Docker containers on a Raspberry Pi. A few are genuinely novel. All of them taught me something.

1. Home Chores — the housekeeper's print list

My housekeeper has been with my family for years. She knows her job, but coordinating recurring tasks — which bathroom gets cleaned on which day, what's on the menu this week, what needs special attention before guests arrive — involves a lot of back-and-forth that neither of us enjoy. I wanted to automate that.

What it does: Home Chores is a Next.js web app that generates daily checklists and prints them — physically, on a thermal printer — for household tasks. It's explicitly not a task tracker — there's no "done" state, no completion logging, no dashboard. The system decides what appears on today's list based on recurrence rules, and the list comes out on paper each morning. The housekeeper grabs the sheet and gets to work.

It also handles meal planning: a monthly calendar where dishes are scheduled (or randomized), so the shopping list and daily cooking instructions flow naturally from the same system.

Home Chores
🖨️ ⚙️
Hoje
terça-feira, 11 de março
Preview do que será impresso para este dia
Joana (diarista) 4 tarefas
🧹Limpar banheiro suíte
🧺Lavar roupa de cama
🍳Preparar almoço: Frango grelhado
📦Organizar armário quarto
4 tarefas regulares · 0 especiais · 1 avulsa
📋
Hoje
Tarefas
🍽️
Cardápio
🖨️
Imprimir
👥
Equipe

The Today view — pulled from a phone at 7am, then handed to the printer.

What's cool: Before this, the housekeeper worked from memory and a pen-and-clipboard system my wife had jury-rigged over the years. Now there's a printed list waiting every morning, and we've been running on it for six weeks without a single coordination hiccup. My wife runs it — she plans the menu, sets special tasks, adjusts. When the AC filter needed cleaning she just added it. Done. It has quietly become the connective tissue of our household operations.

Built in: A weekend, between playing with my kid, barbecuing and jumping in the pool.

2. AIR — the AI news curator

I read Hacker News obsessively. Too obsessively. I wanted a daily digest of only the AI-related stories, summarized, ranked by score, delivered to a static website I control.

What it does: AIR (AI Engineer News) is a fully automated pipeline that runs every morning at 6am UTC. It fetches the previous day's HN stories via the Algolia API, classifies each title as AI-related or not using GPT-4.1-mini, crawls the actual URLs (with Playwright, plus Firecrawl as a paywall fallback), generates ~128-character summaries, and publishes a static HTML page to S3. There's a CloudFront distribution in front of it. Stories with a score above 82 get a fire emoji.

air.lucasprim.org/2026-03-10.html
> AI_ENGINEER.NEWS
// curated from Hacker News with AI · March 10, 2026
<-- prev Monday, March 10 next -->
Significant improvements to long-context reasoning and multi-step tool use; API pricing unchanged from prior generation.
342 pts by throwaway_ai [hn]
Google researchers show LoRA fine-tuning of Gemma 3 27B achieves near-frontier results on HumanEval and SWE-bench.
178 pts by ml_practitioner [hn]
Thread: 312 comments covering Claude, GPT-4.1, open-source models, RAG pipelines, and agent frameworks in production.
95 pts by curious_engineer [hn]
3 of 14 stories shown · ★ = score ≥ 82 · source: Hacker News

AIR's generated newsletter page — dark terminal aesthetic, monospace throughout, stories ranked by HN score.

What's cool: I built something similar in 2023. It worked, barely, and took weeks of wrestling with scrapers, rate limits, and brittle parsers. Claude built a better version — multi-layer crawl strategy, parallel workers, exponential backoff, paywall detection — in a single sitting. It runs unattended every morning at 6am UTC and I haven't touched it since. Live at ainews.lucasprim.org.

Built in: A single sitting.

3. Let's Stretch — stretch reminders for desk workers

I sit at a desk for too many hours. My back knows it. My physiotherapist knows it. I needed something that would interrupt me every hour with a specific stretch, walk me through it step-by-step, and then go away.

What it does: A native macOS menu bar app. 18 curated stretches designed for desk workers. Configurable intervals. Step-by-step visual instructions with a countdown timer. Auto-play sessions that chain multiple stretches. Starts at login, stays out of the way.

🧘
Neck Tilt Stretch
🏃 neck ⏱ 20s 🪑 desk-friendly
Gently releases tension in the neck and upper trapezius muscles.
14
How to do it
1
Sit up straight with your shoulders relaxed and feet flat on the floor.
2
Slowly tilt your head toward your right shoulder until you feel a gentle stretch on the left side.
3
Hold for the duration of the timer, then switch sides.

The Let's Stretch popover — drops from the menu bar, counts down, guides you through each step.

What's cool: This was my first Mac app ever. I've always wanted to ship something native for macOS — it's been a dream since I got my first Mac — but I never learned Swift. I still haven't, really. I watched Claude write the whole thing, stepping in only to decide what the next feature should be. The moment the popover dropped from the menu bar and the timer started counting down, I felt something I haven't felt in software in a long time. I use it every day.

Built in: While working, in parallel with boring meetings. → github.com/lucasprim/lets-stretch

4. Livelo Notifier — catching loyalty point bonuses

Livelo is Brazil's largest loyalty points program, and its partner multipliers change constantly. When a partner goes from 1x to 4x points, there's usually a brief window to earn a lot of points. I was missing these windows because I don't check the site manually.

What it does: A background service that scrapes Livelo's partner page, compares current multipliers against stored history, and sends Telegram notifications when a tracked partner changes. You can manage tracked partners directly through the Telegram bot with inline buttons.

🤖
Livelo Notifier
bot
🔥 Promoção detectada!
iFood mudou de para pontos
⏰ Válido por tempo limitado
🔕 Ignorar 📋 Ver parceiros
14:32
Ver parceiros
14:33 ✓✓
📋 Seus parceiros monitorados
🍔 iFood
✈️ Latam
🛒 Rappi
14:33

Livelo Notifier catches a 4× multiplier window on iFood and pings Telegram immediately.

What's cool: I've wanted this for years. I'm a compulsive buyer of points — I used to check Livelo every few days, anxious about missing a good promotion. Since this went live, that anxiety just... evaporated. Promotions still come and go, but now I find out the moment they happen. When I don't get a notification, I stop thinking about it. The app fixed a behavior I couldn't fix with willpower alone.

Built in: A Saturday afternoon, while my toddler slept.

5. Mercado — automated grocery cart filling

Online grocery shopping at my local supermarket involves too much clicking. For a recurring weekly order, I want to add items to the cart without browsing. Mercado is an attempt to automate that.

What it does: A macOS Electron desktop app that maintains a local product catalog from my local supermarket, normalizes product names and categories using OpenAI, and lets you select items to add to your cart — very quickly, since the catalog is in memory.

Mercado — Delivery Fort
Mercado
🔍 Buscar produtos...
✓ Atualizado
Categorias
Todos (847)
🥩 Carnes (94)
🥦 Verduras (67)
🥛 Laticínios (52)
🍞 Padaria (38)
🧴 Limpeza (71)
Todos os Produtos (847)
🫒
Azeite Extra Virgem 500ml
Gallo
R$ 24,90
🍗
Peito de Frango 1kg
Sadia
R$ 18,99
🥛
Leite Integral UHT 1L
Itambé
R$ 6,49
🍞
Pão de Forma Integral 500g
Wickbold
R$ 8,90
🛒 Carrinho (1 item — R$ 18,99)

Mercado's product grid — OpenAI-normalized names, organized by category, with one-click cart building.

What's cool: My wife is the main user. What used to be 45–60 minutes of clicking through a grocery app is now 5 minutes in Mercado. I didn't overengineer it — no integrations with the menu, no multi-providers, no recurring lists, nothing. Just a local Electron app that does one thing. The fact that it's exactly what she needed and nothing more is the point.

Built in: Two evenings.

6. Mibo Snapshot — 1080p camera frames without port forwarding

I'm building a house. There are cameras at the construction site. The site runs on Starlink — great bandwidth, but you can't open inbound ports, so there's no way to reach the cameras directly from outside. I wanted periodic 1080p snapshots uploaded to S3 automatically, and I wanted them without having to drive over there every time I got curious.

The cameras use a proprietary P2P relay protocol that isn't documented anywhere.

What it does: Connects to the camera through Dahua's proprietary P2P relay infrastructure (the Intelbras cameras are white-label Dahua hardware), opens a PTCP sub-channel, bridges it to a local TCP port, and uses ffmpeg to capture 15 frames from the H.264 main stream — then picks the best one using an ML-based quality scorer. No port forwarding. No local network access required. 1080p.

site-snapshot — python
INFO Starting snapshot service (3 cameras, interval: 15m)
INFO [Camera 1] Connecting via P2P relay... (attempt 1/50)
WARN [Camera 1] Relay connection failed, retrying...
INFO [Camera 1] Connecting via P2P relay... (attempt 2/50)
OK [Camera 1] P2P tunnel established → localhost:18554
INFO [Camera 1] Capturing 15 frames via RTSP...
INFO [Camera 1] Frame scores: [42KB✓ 38KB✗ 51KB✓ 49KB✓ 55KB✓ ...]
OK [Camera 1] Best frame: 55KB (frame 12, no corruption)
OK [Camera 1] → s3://my-site-cameras/Camera 1/20260311141503.jpg
INFO [Camera 2] Connecting via P2P relay... (attempt 1/50)
OK [Camera 2] P2P tunnel established → localhost:18555
OK [Camera 2] → s3://my-site-cameras/Camera 2/20260311141518.jpg
INFO [Camera 3] Connecting via P2P relay... (attempt 1/50)
OK [Camera 3] → s3://my-site-cameras/Camera 3/20260311141531.jpg
OK Snapshot cycle complete in 28.4s · Next run: 14:30:00

site-snapshot running — the P2P relay succeeds on attempt 2, 15 frames captured, best one wins.

What's cool: Claude spent about 30 hours on this. I watched it try and fail dozens of times — wrong connection approach, wrong protocol handshake, wrong frame extraction method. I intervened maybe three times to redirect: "try the P2P relay instead of direct RTSP," "capture multiple frames and pick the best one," "check for corruption before uploading." The result is super reliable. It runs on my home server, handles network flakiness gracefully, and the snapshots land in S3 every 15 minutes without complaint.

Built in: One weekend.

7. Water Tracker — drink water, with your webcam as a witness

I don't drink enough water. My endocrinologist says this. Nobody fixes it. My solution: a macOS menu bar app with a webcam pointed at my desk that detects when I'm actually drinking — and a counter to prove it.

What it does: An overengineered macOS menu bar app. The main feature is the webcam detection: it uses Machine Learning to notice when you lift your bottle and drink, so you don't have to manually log anything. On top of that: configurable bottle sizes, daily goal tracking, screen-edge nudges, and a calibration flow that trains the model on your specific camera, your specific lighting, and your specific bottle.

1 250
/ 2 000 ml
🔥 7 day streak

Water Tracker's menu bar popover — 1,250 of 2,000 ml logged, 7-day streak intact.

What's cool: The webcam detection actually works, which surprised me. The calibration UI — capture baseline, capture drinking, tune thresholds, test live — turned out to be harder to design than the detection algorithm itself. Lighting changes all day and your camera sees everything differently at 3pm vs. 8am. More importantly: I actually drink more water now. Every 15 minutes, my screen flashes blue. That's the reminder. It works.

Built in: One weekend. → github.com/lucasprim/water-tracker

8. Obra Reporter — construction site surveillance via Telegram and Claude

I'm building a house. The construction site is a few kilometers away. I have cameras on-site, their snapshots going to S3 via Mibo Snapshot (see above), and my anxiety keeps tempting me to open the cameras every hour or so.

What it does: A Telegram bot that connects to the S3 bucket where the camera snapshots land. Every hour, it sends an automated progress report — a photo from each camera with a timestamp. At end of day, it generates a summary log. You can also ask it on demand: /fotos shows the latest from each angle, /timelapse compiles the day's snapshots into a video, /analisa runs Claude Vision on the latest photos and tells you what changed.

🏗️
Obra Reporter
bot · online
/fotos
09:14 ✓✓
📸 Últimas fotos — 14:30
Camera 1
Camera 1 2
Camera 2
Camera 3
🎬 Timelapse 🔍 Analisar
09:14

Obra Reporter shows the latest snapshot from each camera — all pulled from S3 on demand.

What's cool: Knowing what's happening at the site is amazing. Every day there are surprises — a new wall up, a team working I didn't know was coming, something unexpected uncovered. My contractor was genuinely taken aback when I showed him the hourly reports and the timelapse from the first week. He asked if I was building something to sell. I said no, it's just for me.

Built in: Two evenings.

9. My Dream Todo App — Obsidian-backed task management

I've tried every task manager. None of them stick. The problem is always the same: the data is locked in the app, the app's model doesn't quite match how I think, and eventually I abandon it. My notes, though — those live in Obsidian, and I've been consistent there for years. So: what if my task manager was just a better interface over my existing Obsidian vault?

What it does: A Next.js web app that reads and writes .md files directly in an Obsidian vault directory. SQLite acts as a derived query index (rebuilt from disk whenever files change), and Chokidar watches for external changes from Obsidian Sync. The source of truth is always the markdown files — which means everything stays in sync with Obsidian on my phone and other devices.

mytasks.local/today
My Tasks
📥 Inbox 4
⭐ Today
Projects
📁 Blog
📁 Stone
📁 Reforma
📅 Calendar
🔍 Search
Available Tasks
📥 Inbox 4
Responder email do contador +
Revisar métricas Stone +
Comprar passagem viagem +
Today's Plan 2 tasks · 1 event
1 of 2 done · 50%
Calendar
1:1 com o time de produto 🕐 14:00–14:30
Publicar post no blog
Revisar métricas Stone

My Dream Todo's planning view — drag tasks from inbox (left) into today's plan (right). Source of truth stays in Obsidian .md files.

What's cool: Source of truth is always the .md files — not the database, not the app's state. Every change I make in the web app syncs automatically to Obsidian on my phone. Every note I take on my phone shows up in the web app. The database is just a derived index that can be thrown away and rebuilt. That's the architecture I've always wanted from a task manager.

Built in: A few evenings.

10. Stone Trainer — practice selling to an AI persona

Sales training at scale is expensive. Role-play sessions with a coach or a colleague are limited by calendar and courage — nobody likes to be the one who stumbles through an objection in front of their manager. What if a sales rep could practice with an AI persona that pushes back, forgets nothing, scores every call, and is available at 11pm on a Sunday?

What it does: Stone Trainer is a sales role-play platform where reps practice conversations with AI-powered customer personas. Each persona has a name, backstory, company, difficulty level, and a set of objections appropriate to their profile. Reps can practice via text chat or voice — and the voice mode is genuinely impressive: sub-300ms latency, natural turn-taking, and a live transcript that follows the conversation. After each session, an AI evaluator scores the call against a structured rubric and generates detailed feedback. There are learning tracks with progressive difficulty, a leaderboard, and certificates when tracks are completed.

Stone Trainer
📊
🏆
🎙️
💬
📋
Trilha: Vendas Enterprise · Estágio 3 de 5
👩‍💼
Carolina Mendes
Diretora de TI · Grupo Financeiro ABC
IA falando...
TRANSCRIÇÃO
Você: Nossa plataforma reduz o custo de atendimento em até 40% no primeiro trimestre.
Carolina: Quarenta por cento é uma promessa bem ambiciosa. Quais empresas do setor financeiro vocês já atenderam com esses resultados?
Objetivo da reunião
Apresente a plataforma e convença Carolina a agendar uma PoC de 30 dias com a equipe técnica.
Critérios avaliados
Abertura e rapport
Descoberta de dores
Proposta de valor
Tratamento de objeções
Fechamento
Dificuldade
DIFÍCIL

Stone Trainer's voice call interface — a real conversation with an AI persona, sub-300ms latency via WebRTC + OpenAI Realtime API.

What's cool: Our operations teams had been asking for something like this for years. Sales training is expensive, scheduling is hard, and nobody wants to stumble through objections in front of their manager. The idea had been on a whiteboard for a while. I sat down with Claude Code and had a working prototype — voice included, ~300ms latency, AI persona responding, scoring rubric evaluating — in under 24 hours. From idea to something you could demo to a sales team: one day.

Built in: Less than 24 hours.

11. Stone AI Apps — AI micro-apps for Brazilian entrepreneurs

Small business owners in Brazil — restaurant owners, barbers, boutique operators, gym managers — are drowning in operational complexity. Marketing, suppliers, customers, products, employees: there's always more to manage than there are hours in a day. Stone serves millions of these businesses with payments and financial services. We'd always wanted to help them beyond banking. But "help" is vague. What does it actually look like? We'd been talking about it for months without shipping anything.

I built the first version on a taxi on the way to a meeting in São Paulo. A team took it from there. In less than a week it was being tested by real customers. The feedback was immediate and genuinely surprising.

What it does: Stone AI Apps is a platform of mobile-first micro-apps powered by AI to help our customers — restaurants, gyms, boutiques, you name it — handle their day-to-day operations. There are apps for photo enhancement, social media content, menu translation, competitor intelligence. There are new ones being built as fast as we can think of them. Each one tackles a specific operational pain. Each one is available to any Stone customer from day one.

aiapps.internal.example.com/dashboard/apps/photo-enhance
Stone AI Apps
🏠 Dashboard
📸 Melhorar Foto
✨ Post Instagram
🌐 Traduzir Cardápio
👁️ Intel Concorrente
SM
Sabor & Cia
Restaurante
Melhorador de Fotos
Transforme fotos comuns em imagens profissionais com IA
Variantes geradas — Prato de Frango Grelhado
🍗
Studio
Fundo branco
🍽️
Lifestyle
Em uso
Gerando...
Contexto
Ambiente
Aguardando...
Criativo
Artístico

The Photo Enhancer streaming four professional variants in real time via SSE — studio and lifestyle arrived, context is generating, creative is queued.

What's cool: AI gave us the ability to jump from "we should help our customers with more than banking" to actually doing it. That gap — between the idealized version of a thing and the real thing in front of real users — used to take months. Requirements, design sprints, estimates, dev cycles, testing. Now it takes days. The cycle compressed so much it's almost unrecognizable. And the apps keep coming.

Built in: A taxi trip and a daylong meeting.

12. Lucy Manager — the migration that never happened, done in two hours

Every engineering team has one. The migration that's been on the roadmap for three years. The one that keeps getting pushed because there's always something more urgent. Ours was Lucy Manager: the internal tool our team uses daily to configure bots, manage knowledge bases, and debug conversations — sitting on a Vue.js codebase that nobody on the team uses anymore. Everyone writes React. The app worked fine. Migrating it "just" meant rebuilding the whole thing in Next.js. Nobody ever found the time.

On a Friday afternoon, a colleague and I sat down with Claude Code at 5pm. At 7pm, we had a working Next.js app.

What it does: Lucy Manager is the control plane for Stone's Lucy platform — our LLM chatbot infrastructure. From it, you create and configure bots (engine selection, personas, settings), manage the knowledge base (documents organized in folders, with draft/published versioning), define conversation intents (seven types: RAG, Dialogflow, FormFlow, GuidedFlow, and more), simulate conversations against draft or live content, and audit historical conversations with feedback, debug data, and audio playback. There's also an embedded chat widget with multiple authentication strategies for different deployment contexts.

lucy-manager.internal.example.com/bots
L
🤖
📄
💬
🔍
Bot: Lucy Atendimento
Rascunho
Publicado
Bots
4 bots configurados
🤖
Lucy Atendimento
GPT-4.1 · ~150 documentos · ~20 intents
ATIVO
🤖
Lucy Vendas
GPT-4.1 · ~70 documentos · ~10 intents
ATIVO
🤖
Lucy Cobrança
Claude Sonnet 4.6 · ~30 documentos · ~10 intents
RASCUNHO

Lucy Manager's bot list — the central control panel for Stone's chatbot infrastructure, now in Next.js.

What's cool: This had been on the roadmap for three years. Every quarter someone would say "we really should migrate Lucy Manager" and every quarter something more urgent would take the slot. At one point we tried to outsource it — the quote came back at six months and R$1 million. We passed. Two hours with Claude Code on a Friday evening and it was done. Our team saw it Monday morning and couldn't believe it. The migration that never finds the time turns out to take an afternoon when you stop writing the code yourself.

Built in: Friday, 5pm to 7pm, with a colleague.


The pattern I noticed

Looking at this list, a few things stand out.

First: the scope of what I was willing to attempt changed completely. Mibo Snapshot required reverse-engineering a proprietary UDP-over-relay protocol, building an image quality scoring heuristic, and orchestrating ffmpeg through an async Python process. A year ago, I would have looked at that problem and said "too much unknown territory for a weekend project." With AI, I treated it as a normal weekend project. And it worked.

Second: the bottleneck shifted from implementation to specification. The hardest parts of every project were figuring out exactly what I wanted — the product decision, the UX model, the right abstraction. The actual writing of code became almost secondary. Home Chores taught me this most directly: the insight that it should be a print list generator, not a task tracker, was the real work. Implementing it was just... typing.

Third: the native platform wasn't a barrier anymore. Let's Stretch and Water Tracker are both native Swift macOS apps with full App Store-quality UX. I had never written Swift. That didn't matter. The language barrier collapsed. What used to require dedicated iOS/macOS engineers — or months of re-learning — is now something you can pick up for a weekend project.

Fourth: the gap between "idea" and "thing that exists" is now measured in hours, not weeks. Every single project on this list, from conception to first running version, took less than a weekend. Some took a single night. The Livelo Notifier, including the Telegram bot integration, the database schema, and the scraper, was probably 4 hours of actual elapsed time.

The premises on which the entire software development lifecycle was built — scarcity of engineering time, high cost of exploration, the need for specialized skills — have changed overnight.


What this means for the SDLC

Our entire field has been organized around a fundamental constraint: software is expensive to build. That constraint shaped everything. Why we have product managers (to prioritize ruthlessly because engineering time is scarce). Why we have sprints (to batch work into predictable delivery units). Why we have separate design, frontend, backend, and QA roles (because expertise is narrow and specialists are faster than generalists). Why we have architecture review boards and RFC processes (because the cost of a wrong architectural decision is enormous to unwind).

All of that organizational machinery makes sense if, and only if, the underlying constraint holds: that engineering time is genuinely scarce and expensive, and that mistakes are costly to fix.

What I experienced over the last six weeks suggests that constraint is weakening. Fast.

When one person can build twelve pieces of software in six weeks of evenings — software that spans mobile native apps, background services, sales training platforms, Electron apps, Next.js web apps, Python data pipelines, and protocol reverse-engineering — the math changes. You don't need to prioritize as ruthlessly if exploration is cheap. You don't need specialists if the language barrier has collapsed. You don't need formal RFC processes if a wrong architectural decision takes an afternoon to unwind rather than a quarter.

This doesn't mean engineers become irrelevant. It means the job description changes. The most valuable thing an engineer does is understand the problem deeply enough to know what to build. That judgment — "this should be a print list generator, not a task tracker," or "our customers don't need better marketing tools, they need help with the messy whole of running a business" — is not something AI does easily. It comes from years of domain knowledge, from real conversations with real customers, from the instincts built by getting things wrong. The kind of insight that produced Stone AI Apps didn't come from a prompt. It came from three years of being close to the problem. AI is remarkably good at the part that comes after that judgment is made.

The SDLC was designed for a world where implementation was the bottleneck. We built enormous systems of process, coordination, and governance to manage that bottleneck. Those systems worked. They delivered results at Stone; I've seen what they can do at scale.

But the bottleneck is moving. And organizations that adapt — that rethink what product management looks like when exploration is cheap, that rethink what engineering teams look like when one person can do what used to require five, that rethink what "done" means when you can ship a prototype in a day — those organizations are going to eat the ones that don't.

I fell from pole position to last place in a single session with Claude Code. Six weeks later, I feel like I've closed the gap. Not because I've figured out how to use AI in my workflow. Because I've started to understand what the workflow itself needs to become.

The renaissance is just getting started.

← All posts