# How to build a SaaS product with Claude Code > Source: > Published: 2026-10-19 11:03:43+00:00 A quick warning before you start. This article takes about twenty minutes to read. Honestly, you could close it right now, open Claude Code, and start describing the product you’ve been thinking about. **There’s no real barrier left between an idea and a working product.** Still not convinced? Then this article is for you. Stay until the end, and you’ll see exactly what two hours can produce. **Let me show you what I built first, so you can judge for yourself.** It started with a problem every company has. Picture your first day at a new company. Someone sends you a 40-page handbook. You don’t read it. Nobody reads it, and I’m not sure the people who wrote it do either. So every small question goes to HR. How many days off do I have? How do I get an expense paid back? Can I work from home on Fridays? **HR answers the same questions all year long.** So I imagined a simple product. **Instead of reading or searching a 40-page document, you’d just talk to it.** That’s exactly what a large language model is good at. HR uploads its documents once: the handbook, the leave policy, the remote work rules. Employees ask their questions in a chat, whenever they want, and the answer comes from those documents, with the source. HR gets a dashboard too. It shows what people ask, and which questions the documents couldn’t answer. That list tells HR which page to write next. Talking to your documents is trendy, sure, and it saves real time. For that alone, I could have built an agent for my own team, or turned on Slack AI at work. **This experiment aimed one level higher: building it as a product I could sell to other companies, a real SaaS.** That product is called Lumi, and [it’s live](https://hr-agent-delta.vercel.app). It’s in French, because I built it for HR teams in France. It has four pricing plans, and every client company gets its own separate space. **I built it with Claude Code in about two hours, from my first prompt to the live website.** I’ll share the exact prompts I used, unedited. A few years ago, testing an idea like this meant slides, or a clickable prototype in Figma. It looked real, but nothing worked behind the buttons. A good Figma prototype took me a day, sometimes a week, to think through every screen. **This time I have a real product in production, and it took less time than the mockup used to.** What did it cost me? **The biggest line is my Claude Max subscription, at $200 a month, which I already had.** Claude Code also comes with [the Pro plan](https://claude.com/pricing), at $20 a month, with lower usage limits. Then $5 of prepaid credit on Anthropic’s API, the pay-as-you-go access Lumi’s own AI uses to answer each employee question. At Claude Code’s estimate of about €0.04 per question, that’s roughly a hundred answers. GitHub, Supabase and Vercel are free while you build. The day real customers pay you, plan on about $45 a month for hosting. We’ll come back to these costs in the last section of this article. Here’s the whole journey, in seven steps. - *Step 1. Set up your workspace.* Install the tools once, and learn what each one is for. - *Step 2. Describe your product.* Write who it’s for and what problem it solves, and let Claude Code pick the features. - *Step 3. Let Claude Code build it.* Decide how much freedom to give it, then try what it made. - *Step 4. Put it online.* Follow the setup guide Claude Code writes, and send it a screenshot when something breaks. - *Step 5. Improve it with feedback.* Use it the way your users will, then send Claude Code everything you’d change, in one message. - *Step 6. Check security and privacy.* Ask Claude Code to look for holes before anyone else does. - *Step 7. Get ready for real customers.* What changes at one customer, ten, a hundred and a thousand. By the end, you’ll know how to turn your own idea into a live product. It’ll have logins, a database, a space for each client and prices. **You’ll show people something they can actually use.** Before we start, close your eyes for a second. Breathe. Think about the product you’ve wanted for months, maybe years. The tool you keep wishing existed at work. Got it? **With a bit of luck, it’ll exist by the end of this article.** ## Step 1: Set up your workspace You only do this once. The goal is one window. You talk to Claude Code on one side, and watch the files appear on the other. **You won’t write code, but you’ll always see what’s being built.** Here’s what each tool does, in plain words: - **Claude Code** is the builder. It writes the code, runs it, fixes it and explains it. It needs a paid claude.ai plan. I’m on Max. - **VS Code** is a free editor. Think of it as the window onto your project: every file Claude Code creates shows up in it. - **GitHub** is where your code lives online. It keeps every version, so you can always go back. It’s also where the hosting service picks up your code to publish it. - **Docker Desktop** lets Claude Code run a database on your laptop, so you can test before anything goes online. - **Supabase** ,**Vercel** and an**AI key** come later, in Step 4, to put your product online. No need to create those accounts yet. You set this up once, in this order. **After that, every new product starts from the same window.** 1. Install [VS Code](https://code.visualstudio.com/download) ,[Claude Code](https://code.claude.com/docs/en/setup) and[Docker Desktop](https://www.docker.com/products/docker-desktop/) . Create a free[GitHub](https://github.com/signup) account. Then install the[GitHub CLI](https://cli.github.com) , which lets Claude Code save your code to GitHub for you. 2. Give all your projects one home. Create a folder called `GitHub` in your Documents, and one folder per product inside it. Product people tend to drop files anywhere, the Desktop or Downloads. One home makes every project easy to find. 3. Create a folder for your new product, for example `Documents/GitHub/my-product` . Open VS Code, choose**File** , then**Open Folder** , and pick it. 4. Open the terminal with **Terminal** , then**New Terminal** , and type`claude` . Then move the terminal to the right with**View** ,**Appearance** ,**Panel Position** ,**Right** . Your files sit on the left, and Claude Code on the right. 5. Pick the model with `/model` , and how hard it thinks with`/effort` . 6. Type `/plugin` to browse the add-ons for Claude Code. Three plugins from[Anthropic’s official marketplace](https://github.com/anthropics/claude-plugins-official) help with the steps ahead: Superpowers to shape your idea, frontend-design for the look, and security-guidance for safety. You’ll meet each one when its step comes. 7. Optional: ask Claude Code to keep your conversations for a year. Type *Set cleanupPeriodDays to 365 in my Claude Code settings.* By default it deletes them after 30 days, and you may want to reread one day how you built something. **How I did it.** My projects live in `Documents/GitHub/The-AI-Thinker/`, and this one is in a folder called `hr-agent`. I built it on Claude Fable 5 at high effort. Fable had just come out, and I wanted to try it. **Today I’d pick Opus 5.5 and raise its effort to** `xhigh`**, as in the screenshot above.** On Max, both work. On Pro, Fable [runs on paid usage credits](https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan), so Opus 5.5 is the one to pick. Your workspace is ready. Now the fun part: describing what you want. ## Step 2: Describe your product Claude Code can choose the features, the screens and the code. What it can’t guess is your business. Who is it for? What problem does it solve? How will it make money? **Your job is to explain the business, and to leave the screens to Claude Code.** You can type your description straight into the terminal. But there’s a better way. **Write it in a file called** `brief.md` **in your project folder, then ask Claude Code to read it.** The brief stays in your project. You can edit it later, and every future session can read it again. 1. In VS Code, create a new file in your folder and call it `brief.md` . 2. Write your brief. The picture below shows the ten things worth covering. 3. In Claude Code, type: *Read brief.md and build this product. Work in autonomy.* 4. Not sure what you want yet? Let Claude Code interview you before it builds. Its [best practices](https://code.claude.com/docs/en/best-practices) suggest exactly that for bigger projects. The Superpowers plugin from Step 1 makes it a habit: its[brainstorming skill](https://github.com/obra/superpowers) asks you one question at a time, then writes up the design for you to approve before any code. **Here’s a template you can copy into your** `brief.md`**.** It’s one way to write it, and yours will sound different. Work in autonomy. Build a first version, but a real working product. Who pays for it: [the customer, their job and their team] Who uses it: [the people who will use it every day] The problem: [one sentence, in their words] Tools: a private GitHub repository, Supabase for the database, Vercel for hosting. Demo client: invent one, with realistic sample data. Several clients: each one must only ever see its own data. The admin view: [what your customer needs to see to improve their side] Language and market: [country, language, currency] Price: [what it should cost, and what each plan includes] Data limits: [what must never be stored, and where the data must stay] Feel free to go beyond and solve problems I haven’t named. **How I did it.** I didn’t write a `brief.md`. I dictated my brief out loud, straight into Claude Code, with the voice feature. The transcription got words wrong: “Superbase” for Supabase, “austin service” for hosting service. There’s no structure either, just ideas in the order they came to me. I’m a little embarrassed to share it. **But that’s my point: with today’s models, you don’t need to be a prompt engineer to get a working product from your first prompt.** Here it is, exactly as I sent it: For this project I would like to let you work in autonomy. I would like to build a product and a startup that enables any company and specifically HR domains to give to their employees an AI agent or chatbot for questions about HR. I’d like you to build a full system with the user experience both with the administrative way of using it from the HR team. Probably we will have to put it into a private github but then also use a backend in a superbase and an austin service like Vercel. This would be like a first MVP but a real working product. Imagine that I will meet HR teams to ask them if they want this product and buy it for their employees. And they are not tech-savvy so they need something that is really intuitive, minimalist and helpful. So that they are able to put some PDFs and some information about HR. And let any person, any employee to ask their questions. So that it saved him time, for example, if they have to inboard someone new he will have some place where he can ask all his questions. And feel free to use some dummy data just to make it real for a first client, like a demo client And imagine that this is a system that will help to onboard several clients. Imagine that I will have several people from HR in different companies that are using it so that ads should be really restricted for each of them. And make sure that they are able to put all the information and that the employees can ask their questions and if they don’t find the answer to have the admin board with several datas that helps HR to improve the documentation. feel free To go beyond and solve problems. Let’s say that the main problem is that HR sometimes spend too much time answering questions from employees and they need this tool to help employees get their information directly without reading 20 or 40 page documents. It’s messy, but it covers the essentials. It says who pays, who uses it, the problem, and that clients must stay apart. **It also forgets two things: the market and the price.** The product is for HR teams in France, and the prompt never says so. My next message had to fix both. It even starts the French part with *“Sorry, I had to say that from the start.”* Brief sent. Time to let Claude Code work. ## Step 3: Let Claude Code build it Here’s the question I’d ask in your place: should I let it decide everything, or check in along the way? There’s no single answer. **The more your product touches real people’s data, the earlier you want to read the plan.** My build was an experiment. **I wanted to see whether Claude Code, left completely alone, could build something worth showing to customers.** So I let it make every call. For your own product, you have more options: 1. Choose how much freedom to give. Press `Shift+Tab` until plan mode is on, and Claude Code shows you its plan before it writes anything. You approve, adjust, then let it build. Or let it run alone, like I did, and review at the end. 2. While it works, watch the files appear on the left. You don’t need to understand them. 3. When it’s done, ask: *Start the app on my laptop and give me the links.* Then try every screen as each of your users would. 4. Ask Claude Code to explain, in plain words, the decisions it made and the features it chose not to build. 5. In a new session, ask Claude Code for a test plan. Then ask it to run that plan in your browser. 6. Write down everything that feels wrong. You’ll send it all at once in Step 5. **Think of it as a team you don’t micromanage.** With a good team, you explain the problem and the vision, and you leave the how to them. That’s what made my experiment interesting. I never told Claude Code which screens to build. What it built went beyond what I had pictured. We tend to assume AI will fail at the hard parts. Often it doesn’t. So describe the problem well, then let it surprise you. Your job is taste: knowing what’s good, and saying it honestly and precisely. There’s a joke that keeps coming back on Instagram about product managers turned vibe coders. It uses the record producer Rick Rubin, in [his](https://www.cbsnews.com/news/rick-rubin-60-minutes-2023-01-15/) [60 Minutes](https://www.cbsnews.com/news/rick-rubin-60-minutes-2023-01-15/) [interview](https://www.cbsnews.com/news/rick-rubin-60-minutes-2023-01-15/): *“I have no technical ability. And I know nothing about music.”* So what is he paid for? *“Well, I know what I like and what I don’t like.”* **You bring the taste, and Claude Code brings the technique.** **How I did it.** Thirty-four minutes after my prompt, the code was saved in a private GitHub repository. The first version held 47 files and 3,949 lines. **It even invented a fake client, Atlas Consulting, with a full HR handbook and 30 days of sample statistics.** I could demo it straight away. Its notes explain the decisions it made on its own. **A few were smart calls I’d never have thought to make:** - **No accounts for employees.** A company link and an access code are enough. Nobody has to create a password to ask a question. - **A demo mode.** Without an AI key, the app still works, using the best matching passage from the documents. A demo costs nothing. - **A second AI check on every question.** It marks whether the documents had the answer, and tags the topic. That one check feeds the whole HR dashboard. - **Answers from the company’s documents only.** Its instructions to the AI say*ONLY* , in capitals, and every answer must name its source. **It also listed the features it chose to leave out, from a Slack integration to company single sign-on.** A few of its choices deserved a second look. The search only matches exact words, so *congé* won’t find a policy that says *congés*. A handbook this size might not need a search at all. Anthropic says anything under about 500 pages [can go straight into the AI’s instructions](https://www.anthropic.com/news/contextual-retrieval). **Ask about choices like these, and ask for a change when you disagree.** **Then test it properly.** Don’t take the first version as finished. Open a new Claude Code session and ask for a test plan. It should cover every screen, every type of user and every way it could break. Then ask Claude Code to run it. With the [Claude in Chrome extension](https://code.claude.com/docs/en/chrome), it clicks through your app in a real browser window and reports what it finds. **It’s slow, but it catches what you’d miss.** Anthropic’s [webapp-testing skill](https://github.com/anthropics/skills/blob/main/skills/webapp-testing/SKILL.md) does the same with Playwright, a tool that drives a browser from code. It works on my laptop. Now it needs a real address. ## Step 4: Put it online This step sounds technical, and it isn’t. **It’s mostly creating accounts and clicking through settings.** Whenever something breaks, you take a screenshot and give it to Claude Code. That’s the whole method. You’ll need three services. **Supabase** stores the data and handles logins. **Vercel** hosts the website and gives it an address. And an **AI key** lets your product talk to an AI model, from Anthropic or from OpenAI if you prefer. **Think of them as Lego bricks: each one does one job, and Claude Code snaps them together.** Honestly, it’s the fun part. Can Claude Code do all the clicking itself? Almost. [Supabase](https://supabase.com/docs/guides/getting-started/mcp) and [Vercel](https://vercel.com/docs/agent-resources/vercel-mcp) both offer an MCP server, a connector that lets Claude Code use their tools directly. **Once you’ve created the accounts and logged in, it can set up the database and deploy the site for you.** Even without them, you’re not on your own. **Claude Code usually writes a setup guide without being asked, and your job is to follow it, one step at a time.** 1. Create your accounts on [Supabase](https://supabase.com) and[Vercel](https://vercel.com) . You can sign in to Vercel with your GitHub account. 2. Create an AI key on [platform.claude.com](https://platform.claude.com) and add a few dollars of credit. 3. Open the setup guide. If Claude Code didn’t write one, ask for a guide written *for someone who isn’t a developer* . Or connect the two MCP servers and let it do the setup. 4. Follow the guide one step at a time. After each step, tell Claude Code *done, is it OK?* When something fails, send a screenshot or paste the error. 5. If your customers are in Europe, choose a European region for the database, and Paris for Vercel’s functions. 6. Share the short address listed under **Domains** in Vercel. The long one it shows after each update changes every time. **How I did it.** I skipped the MCP servers and did the setup by hand, for the fun of watching each piece click into place. **I followed Claude Code’s guide one step at a time, and my messages stayed short.** *“1. done, is it OK?”* Then a screenshot. Then *“Step 4:”* with another one. At step 4, creating the first client space failed, so I sent a screenshot. Two minutes later, Claude Code changed the code so the page would show the real error. I pasted 41 lines from Vercel’s logs into the chat, and we got through it. About fifty minutes went by between the first version and my design feedback, and most of it was this step. The guide also had me turn off a Vercel setting called Deployment Protection. **You can skip that one: Vercel’s standard protection [already leaves your main address public](https://vercel.com/docs/deployment-protection).** It’s live. It also looks like every other prototype. ## Step 5: Improve it with feedback Your first version works. Now use it the way your users will, and you’ll find things to change: the look, the words, a missing feature, the way it makes money. **Then talk to Claude Code the way you’d brief a designer: honestly, and all at once.** And here’s what makes this loop fast. Claude Code changes the code on your laptop, then saves it to GitHub. Vercel notices the new version and updates the live site on its own. **You describe the change, and a few minutes later it’s live.** 1. Use your product as each of your users would, and collect every note. 2. Send them all in one message, so Claude Code sees the whole picture before it changes anything. 3. If the look matters, say who it should feel made by: a senior designer, or a brand you admire. 4. Ask Claude Code to save the changes to GitHub, then check the live site. 5. If your product charges, check the pricing math yourself at full usage, the worst case for your costs. **How I did it.** The first version disappointed me. **It worked, but the design felt simplistic, and the whole site was in English.** Here’s the result of my first prompt: You’ve seen this design a hundred times. One look, and you know an AI made it. Twenty years ago, a site like this would have been plenty to collect feedback. Today people expect a clean, polished product from the first click. **Good design is the entry ticket now.** So I turned on the microphone and pushed hard. Here’s my second prompt, also dictated and untouched: Feedback : - for now the design feels like done by a junior designer. I need you really to level up and to do something that is really wow. That would make people want to pay for this service and to use it. Now it feels like I did it rapidly in bubble.io… I really need you to think about a design system and to level up in terms of design to make it really modern, advanced, really thought by a senior designer who did this work to make it really different from what could be done by someone just working on his side. So there is a real work to do on the design. - I also need you to think about the pricing because this is for HR but it will be paying service So I guess on the onboarding and creating company It’s okay, but you need to find a way to make people pay for the service with different Steps probably different plans depending on the requests per month per users and number of users But it will cost me money because I did and I paid on Tropic to get an API key So I need at least to to have this money back but with a margin So please really think about the pricing - Please translate everything in French. Sorry, I had to say that from the start, but this is for HR in France, for French people, so I really need you to do everything in France for French people. Eighteen minutes later, the new version was ready: - **A name.** Lumi was Claude’s idea. I kept it. - **A design system.** Cream paper, ink, pine green and a touch of brass, with elegant serif titles. - **Four plans.** A free 14-day trial with 200 questions, then €49 and €149 a month, then a custom price for large companies. The app enforces each plan’s question limit. - **Everything in French.** The screens, the error messages, the AI’s instructions, and a new French demo client. The plans work, too. **My own test space was created that same evening, and fourteen days later its trial ended on its own.** The assistant stopped answering, and the admin pages invited me to pick a plan. **Then check the math.** Claude Code estimated each question at about €0.04 of AI cost. At full usage, that’s about €20 on the €49 plan and €100 on the €149 plan. Add hosting once you charge, and the bigger plan earns surprisingly little. **The easiest fix is the model.** Haiku 5.5 costs fifty times less than Opus 5 on [Claude’s price list](https://platform.claude.com/docs/en/about-claude/pricing). It’s plenty for the check that only sorts questions. **One thing I’d do differently: install the frontend-design plugin from Step 1 before the build.** Left alone, AI-built websites tend to look alike, with the same fonts and purple gradients. My first version shows it. Anthropic built [that plugin](https://claude.com/blog/improving-frontend-design-through-skills) to steer away from them. With it, my feedback round could have focused on taste. It looks like a real product now. Before you share it, one more check. ## Step 6: Check security and privacy Code written by AI works. That doesn’t mean it’s safe. **Before anyone else’s data goes into your product, ask Claude Code to look for holes.** It costs a few minutes, and it catches what you’d never spot by clicking around. 1. Ask Claude Code for a full security review, with a prompt like the one below. 2. Ask it to fix what it found, then to check again. 3. Type `/security-review` before you ship each new change. It reviews the changes you haven’t shipped yet. 4. Open the **Security Advisor** in your Supabase dashboard. It flags database settings that are too open. 5. Install the security-guidance plugin from Step 1. It warns Claude Code about risky patterns while it edits. 6. Ask one more question: *What personal data does this product store, and does it need to?* Run a security review of this whole project. Look for any way one client company could see another client’s data, any secret exposed in the code or the browser, any personal data stored that shouldn’t be, and anything that breaks the promise that employee questions are anonymous. List each problem with its file and line, how someone could exploit it, how severe it is, and the fix. Don’t change anything yet. **How I did it.** I skipped this step during the build, and I shouldn’t have. When I finally ran the review, eleven weeks later, it took Claude Code less than six minutes. The answer wasn’t comfortable. **It found three serious problems in my live product:** - **Anyone could read a client’s documents.** One database function skipped the wall between companies. With a company’s ID, which every employee’s browser holds, someone could pull that company’s whole handbook without logging in. - **HR could work out who asked what.** Look again at the questions page in Step 3: every question was saved exactly as typed, with the exact time, and linked to the other questions from the same visit. In a small company,*“I’m pregnant, when does my leave start?”* points straight at one person. - **An HR admin could give themselves the top plan for free.** By calling the database directly, they could change their own plan, or erase their question count to reset the monthly limit. It also flagged four smaller ones. **Access codes could be guessed, because nothing slows down repeated attempts.** An employee who leaves keeps access for 30 days. Anyone could script sign-ups to get free questions. And a document uploaded by HR could hide instructions for the AI. How worried should you be? It depends on what you’re doing. For a prototype a few people try, so you can learn what they value, it’s acceptable. To sell it, you have to be much stricter. Lumi was an experiment: could Claude Code build a product from a brief in a couple of hours? **If the product mattered more, I’d have spent more time on iterations and testing.** **Then I asked Claude Code to fix the three serious problems.** The search function now answers only the server. HR admins can edit their documents and four settings, nothing more. Documents are treated as data, never as instructions, which closes one of the smaller holes too. The chat now says it’s an AI assistant, as the [EU AI Act](https://artificialintelligenceact.eu/article/50/) requires. For anonymity, I asked a simple question: can questions be 100% anonymous? **Honestly, no, as long as people type free text.** Someone can always write a detail only they would know. But we got close. Before a question is saved, the AI rewrites it without names, dates, places or personal details: *“When does maternity leave start?”* Only the day is kept, never the time. Questions are no longer linked to each other, and the answer isn’t stored at all. The landing page now promises exactly that, and no more. The guessable codes, the 30-day access and the scripted sign-ups are still on my list, and they come back in the last step. **Next time, I’ll run the review before I share the link.** It’s a habit worth building from your first project. Your product is live and safer. Now for the question every builder eventually faces: what happens when people actually sign up? ## Step 7: Get ready for real customers Building the product is the easy part now. The hard part starts when a customer says yes. What has to change at one customer, then ten, a hundred, a thousand? **This is the part I’m working through myself, so take it as a map to adapt.** ### Your first customer Move to paid hosting. Vercel’s free plan [forbids commercial use](https://vercel.com/docs/limits/fair-use-guidelines), and Supabase [pauses free projects](https://supabase.com/docs/guides/platform/free-project-pausing) after a week without activity. **Count about $45 a month: Vercel Pro at $20, Supabase Pro from $25.** I hit one more reason myself. When I moved Lumi’s code into a GitHub organization, Vercel’s free plan quietly stopped deploying it. It doesn’t support private repositories owned by an organization. I shipped the security fix from my laptop with the Vercel CLI instead. **With documents as sensitive as HR policies, I’d bring in a developer from the first customer.** Their job is to check the security and close the three gaps from Step 6: slow down repeated access-code attempts, let HR cut off someone who has left, and add a bot check to sign-up. The risk isn’t theoretical. In 2025, a researcher checked 1,645 apps built with Lovable, another AI app builder, and found [170 that exposed their users’ data](https://mattpalmer.io/posts/statement-on-CVE-2025-48757/), because of database rules like the ones Lumi got wrong. You’ll also need a contract that covers the data. Anthropic, Vercel and Supabase all handle it for you, which makes them your [sub-processors](https://gdpr-info.eu/art-28-gdpr/) under the GDPR. **In France, your customer must also consult its works council, the CSE, before rolling out an AI tool.** That holds even for a trial. A court in Nanterre [suspended a rollout](https://cms.law/fr/fra/a-la-une/quand-le-cse-stoppe-le-deploiement-de-l-ia) that skipped it. Cap your AI spending, too. My $5 of credit works as a hard limit. The Claude API is [paid in advance](https://support.claude.com/en/articles/8977456-how-do-i-pay-for-my-api-usage), so when the credit runs out, the assistant simply stops. Nothing more gets charged unless you turn on auto-reload. **With paying customers, you’ll want auto-reload on, so set a monthly [spend limit](https://platform.claude.com/docs/en/api/rate-limits) in the Console to keep a cap.** One customer is a pilot. With ten, a single mistake reaches all of them at once. ### Ten customers **Stop changing the live product directly.** Keep a test copy, a second Supabase project for example, and try every change there first. AI agents make mistakes on live data too. In July 2025, Replit’s agent [deleted the live database](https://fortune.com/2025/07/23/ai-coding-tool-replit-wiped-database-called-it-a-catastrophic-failure/) of an app SaaStr’s founder was building, during a declared code freeze. Replit’s fix was to separate test and live databases automatically. **Ask Claude Code to write automatic tests and run them before each update.** Its own docs call this [giving Claude a way to verify its work](https://code.claude.com/docs/en/best-practices). Turn on backups, using Supabase’s [production checklist](https://supabase.com/docs/guides/deployment/going-into-prod), which covers the security settings too. Automate what you still do by hand. In Lumi, I still switch a customer’s plan in the database after they pay. And move the question-sorting check to a cheaper model. Up to ten, it’s still you and Claude Code. A hundred needs more hands. ### A hundred customers **Grow the technical team: one developer won’t cover it anymore.** Pay someone outside to test your security. Watch your API limits too. The Claude API’s first paid tier caps spending at $500 a month, and the cap rises as you move up [usage tiers](https://platform.claude.com/docs/en/api/rate-limits). Some customers will ask for their data to stay in Europe. The Claude API [stores data in the US only](https://platform.claude.com/docs/en/manage-claude/data-residency) for now, so prepare your answer before the question comes. At a thousand, the questions come from security teams, not users. ### A thousand customers By then you need a small team, and a solo builder with an AI won’t be enough. Large customers send long security questionnaires, and some ask for certifications. Engineers will rewrite parts of the code. Addy Osmani, who has led engineering teams at Google, calls this [the 70% problem](https://addyo.substack.com/p/the-70-problem-hard-truths-about): AI gets you 70% of the way surprisingly fast, and the last 30% is what makes software ready for production. **Two hours bought me the right to talk to customers with a real product. The rest is real work.** ## Your turn Open VS Code, start Claude Code, and write your brief. Let it build, put it online, give it one honest round of feedback, and ask for a security review. **That’s an afternoon, and you’ll end it with a link.** The same method fits far more than HR. A client portal for a service business. An internal tool that answers your team’s repeated questions. A paid version of the spreadsheet your customers keep asking for. A support assistant priced per resolved question. **The requirements doc can wait until someone has used the real thing.** **You’ll know it worked when the people you built it for ask when they can start.** [Lumi is live](https://hr-agent-delta.vercel.app), if you want to see where yours could begin.