Ship Solo, Scale Together
Speed is easy. Velocity is a team sport
Intro
Speed. That’s what AI tools have brought into our work lives. But raw speed without direction is just noise.
The right move is turning speed into velocity, same force, but now with a clear vector. And that shift is what let me, a single engineer, build two production-ready systems in a couple of months at a 16-year-old company.
This post is about how I did it: the workflow, the technical decisions, and how Alan, Fiverr’s internal AI assistant packed with company context and practical skills, became my reliable engineering partner. If you’ve ever wondered what it actually looks like to ship what once required a full team, this is that story.
Past Trials
AI is moving fast, and leading companies like Fiverr want to react to market changes even faster. This project was a chance to do exactly that: a product that lets Fiverr embed itself into other websites and deliver a full end-to-end flow, where an AI assistant welcomes the user, understands their intent, finds the right freelancer, and handles the full transaction.
The first version was built inside Fiverr’s main codebase, and it was not a regular Fiverr feature. A product that lives on other companies’ websites asks for a different approach from one that lives on our own.
We shipped it. And more importantly, it taught us exactly what we’d do differently next time.
Flexibility of Tech
The second version came with less baggage, but demanded more creativity. Now the product can be built outside of Fiverr’s ecosystem. How would you build a system if you could choose any technology available? How do you use that freedom when there are no tech restrictions at all to deliver two complex products within a limited timeframe? And I don’t just mean framework choice. I’m talking about deployments, build tools, CI/CD, monorepo management. There are so many decisions to make before a project can reach real users.
Get Mert Vasit’s stories in your inbox
Join Medium for free to get updates from this writer.
Oh, and one more thing: this was no longer a team effort. It was just me plus my trusted friend, Claude Code (CC).
A One-Man Band
In this new era, communication skills matter more than ever. Even in human-to-human communication, we rarely express our intentions perfectly. So how do you convey intent clearly to a machine?
What really helped me wasn’t just the back-and-forth itself. I already owned the full context: the backend contracts, the Figma designs, the product requirements. There was no PM to chase for clarification, no handoff gaps. So when I forced CC to ask clarifying questions until we reached a shared understanding, I could answer quickly and precisely. Then we’d lock in a feature roadmap together before writing a single line. The planning cycle isn’t new, but having one person hold all the context made it sharper and faster than any team handoff I’ve been part of.
In practical terms, my job was to build a new FE interface by understanding an existing backend (BE) application and implementing designs from Figma. It sounds straightforward until the user flow gets complex and the room for error grows quickly.
Thanks to the Figma MCP (Model Context Protocol) and Fiverr’s internal design system skill, building the UI felt natural. Fiverr’s design system skill wires Claude Code directly to our component library, giving CC the context it needs to use the right components rather than invent its own. The result: create a roadmap per feature, provide the Figma design link, then sit back and review ☕.
Once the server is connected, the hardest bit is over. The rest is easy: just give CC the Figma link, add a clear prompt, and watch it work. The Figma MCP grabs a design screenshot and the main thing is, I always make sure the final output perfectly matches the exact styles the designers set.
The harder part was wiring in the logic layer against the existing BE. Before any implementation, I made sure Claude reviewed the codebase with my instructions, listed all API contracts, WebSocket event types plus payloads, and summarized the data flow. Only then moved on to the FE design.
Obstacles Along the Way
So far I’ve covered mostly the process side. But technically, things got even more interesting.
The SDK needed to work on any website without inheriting that site’s styles or scripts, like a sandboxed app running inside someone else’s page. I solved that with Shadow DOM, a browser feature that wraps the SDK in an isolated environment. Clean, predictable, no interference. The catch is that isolation also blocks third-party tools like Stripe, which needs to inject its payment inputs directly into the page. The fix was Slots, a way to punch a controlled hole in the isolation, letting Stripe do its thing while the SDK still treats those elements as its own.
The final stack (for the curious): an Nx-managed monorepo with a Next.js dashboard, a React-based UI SDK built with Vite, and shared component libraries. CI/CD runs on CircleCI, with the SDK uploaded to S3 and the dashboard deployed to AWS Amplify.
Finale
All the AI tools that help us move faster truly shine when you direct them with clear intent and the right context. The speed is the easy part, it’s the clarity you bring that makes the difference. But what I built here scaled one engineer, not a team. The real next step is making these patterns shared: skills that encode how your company builds things, context that any engineer can hand to an agent and get consistent results. That’s what tools like Alan, Fiverr’s internal agentic platform, are moving toward and where the real multiplier is. One engineer shipping like a team is impressive. A whole team working that way is a different game entirely.
