I'm the CPO of a fintech company. Me and my technical co-founder shipped a full payment platform in 90 days. Today a year after, we keep shipping new features at a rapid pace. This is the exact playbook we use.
The context
Hi, I'm David Balsam, Co-Founder and CPO of Adesco. In the past year, two people shipped a full multi-tenant payment platform, a backoffice, third-party APIs, and regulatory compliance. No handoffs. No waiting. No lost momentum.
The old way: I had ideas, five people touched them, users saw them six months later. The new way: my co-founder owns the code, security, and infrastructure. I own product, design, and the customer. We both work in the same codebase, each adding where we're strongest. Everything below is the playbook.
I built our early product in Lovable. When we needed to scale, my co-founder rebuilt the architecture in Cursor. At that point he said: why do I rebuild what you can build? He showed me our repo, Supabase, how to branch, how to test on dev. Then he stepped back. I never handed off again.
01 · The two roles
Every team shipping with AI today divides into two roles. These are not titles. They are ways of working. There are architects, they are in charge of the foundations and security of the codebase. The Builder owns the customer experience of the product. AI is what makes the two of us move fast and like we are a fifteen person team.
Product · Design · Customer
Foundations · Security · Infra
The Architect sets the foundations. Before the Builder builds anything.
What is the Builder's job?
Before AI: Customer to PM to Designer to Developer to QA. Two weeks. Translation loses things at every handoff. Today: Customer talks to the Builder directly. The Builder builds the feature that same night. No chain. No delay.
The Builder plans the feature, reviews it with AI, builds it in stages based on the foundations, and tests the result as a user. Every feature passes through these four stages without exception: Plan, Review, Build, Test. No shortcuts.
Screenshot a UI you like and prompt: "Copy this and fit our format." Use shadcn/ui for components. Use 21st.dev for blocks that ship as ready code. The Builder owns consistency across the whole product. No dedicated designer needed.
Marketing page, product, back office, client email, dashboard, CRM. The Builder is the only person who sees the whole picture. Own that view. Make every handoff invisible to the user. And manages the relationships with the stakeholders.
The customer relationship shift
02 · The Builder's mindset
The mistakes I made when I started weren't technical. They were about how I was thinking about the work. Here's what changed.
Your job isn't to write code. It's to write the brief. Structure every prompt like a spec: what you're building, who it's for, what it should do, what it should look like, what states to handle, and what the constraints are. If you can't explain it clearly, the AI won't build it correctly.
AI doesn't plan ahead unless you do. If you just say "build this," it starts at the top and skips everything else. Prompt to get the structure first, make edits, then implement. Every time I skipped this step, I paid for it later.
Default, Loading, Success, Error, Empty. If you only build the success state, you've built half a feature. The other half is what the user sees when something goes wrong. And they always notice.
Don't keep prompting into a broken context. Stop. Describe the problem as a user, not as a developer. Get AI feedback on why it's stuck. Start fresh from the plan. This is almost always faster than pushing through.
Every pattern you build (an auth flow, a file upload, a multi-tenant structure) can be replicated. Turn your patterns into reusable prompts. What took two days the first time takes two hours the fifth time.
Log in. Log out. Go through the whole flow on mobile. Pretend it's your first time. Then try to break it: refresh mid-form, submit strange data, navigate away early. If it handles all of that, you're done.
03 · The prompt formula
If you can't explain it clearly, the AI won't build it correctly. These six elements work at the beginning of a project and at the feature level as you go.
04 · The stack
I tried a lot of tools. These five stayed. Each one does one thing well, and they connect cleanly.
My main coding environment. I write every feature here using AI agents. It reads my codebase, understands context, and follows my custom skills.
Primary IDEAnthropic's CLI that runs directly in your terminal. It reads your entire codebase, executes commands, edits files, and runs tests. Best for complex multi-step tasks where you need an agent that actually understands context across files.
Agentic CLICopy-paste UI components built on Tailwind and Radix. You own the code, not a library. Buttons, modals, tables, forms. All consistent, accessible, and easy for Cursor to extend because the code is right there in your project.
UI componentsThe entire backend: auth, database, file storage, edge functions, and Row-Level Security. It's what AWS would be if someone actually designed it for people who aren't backend engineers.
Version control and your safety net. Branch for every feature. Commit often. If an AI change breaks something, you go back. If an experiment fails, you go back. Non-negotiable.
Undo button for everything05 · Cursor skills
A Cursor skill is a saved instruction set. You run it once, get a structured output, and act on it. Every skill below is free to copy from GitHub. Each one replaces something I used to wait on — a review, a checklist, an update email.
Point it at any page. Returns a letter grade (A to D), a line-by-line critique of what's killing conversion, and a ranked fix list you can act on in under an hour. I run this before every landing page goes live.
Reads every sentence and flags what's vague, off-brand, or too long. Returns specific rewrites, not just direction. Run it on any page, email, or blog post before it goes anywhere a customer might read it.
Runs 12 checks across meta tags, heading structure, internal links, and AI-readability signals. Returns pass/fail on each with a specific fix for every failure. Important now that ChatGPT and Perplexity cite pages directly.
Maps friction points by severity. Catches missing feature states (what shows when the list is empty?), accessibility gaps, and mobile issues your eye misses on a desktop screen. Run it after every new UI component.
Reads every file that changed before a deploy. Returns GO, NEEDS FIXES, or DO NOT DEPLOY with a specific list of what to address. Built for fintech: catches floating-point money bugs, missing RLS, and cross-app breakage. Runs in about 20 seconds.
Reads your git history and writes a plain-English summary of what shipped. Safe to paste into Slack, drop into a client email, or send to your board. No jargon, no commit hashes. Just what changed and who it helps.
Bring this to your team
This playbook is not just for product teams at startups. I've taken it to companies across Israel and taught their product, tech, and leadership teams how to move faster with AI. If you want me to run a session at your company, fill in your details below and I'll get back to you.