Shipping faster with AI, without losing the craft.

I build with AI every day, on Scenario Builder at TransitionZero and on my own products. It’s changed how fast I get from idea to production. I still own the architecture, the UX detail and the quality bar, and more and more of my work is building the systems that let agents meet that bar. Here’s how that works in practice.

How I work

  • Specs before code

    Features start as a spec, worked out by having an agent grill me until the edge cases are on the table. The spec becomes the plan, the issues and the record of why.

  • Reference prototypes

    I take a spec straight to a high-fidelity, working prototype, sometimes pairing with a designer, so we can test whole user journeys end to end. With AI, prototype code is cheap, so the prototype becomes the reference for the production design and build.

  • Parallel agents, layered review

    Several Claude Code sessions run in parallel, test-first where it fits. Every change gets an AI first-pass review before I read it.

  • Codebases built for agents

    Something I took from Lauren Tan: to handle thousands of PRs a month, a codebase needs strong architecture and static analysis first. On Scenario Builder I’m building custom skills, agents, rules and strict static analysis so that backend engineers can ship frontend features agentically with little or no input from me.

  • Trust, verification and quality gates

    If I review every line myself, I become the bottleneck. On A Legacy Beyond, Claude writes the code, CodeRabbit and GPT review it, and I audit those reviews against coding standards I set up at the start. Trust comes from verification, and that’s what lets a software factory scale to dozens of agents.

  • A software factory

    Beyond engineering, I’m working with product, design, and AI and data specialists to build a quality-controlled software factory around these agents.

Tools

  • Claude Code

    My main tool for building, refactoring and reviewing. Remote Control lets me keep agents moving from my phone.

  • Skills

    Matt Pocock’s skills are my base, with Superpowers for brainstorming. On top sits my own suite of custom skills, some written from scratch and some adapted from popular ones like Grill Me.

  • MCP servers

    Figma, GitHub, Postman and Vercel, among many others. MCP gives agents the same context I have, from designs and issues to APIs and deployments, so I’m not pasting it in by hand.

  • Claude Design and Artifacts

    For exploring designs and visualising flows as diagrams.

  • Verification

    Claude in Chrome and Chrome DevTools let agents check their own work in a real browser, profiling included.

  • Second opinions

    ChatGPT when I want another model’s take, and CodeRabbit in code review.

  • Building AI products

    AI SDK, AI Gateway and AI Elements for customer-facing AI features, and Vercel Eve when a product needs a full agent framework.

Learning

  • MEGA

    I’m in the first cohort of mega.dev’s program for developers building with agents, taught by Kent C. Dodds, Angie Jones, John Lindquist and Theo Browne.

  • What’s next

    Going from orchestrating 1–5 agents to 10+. Today I lean on structured, hand-holding workflows. Next I want enough trust in them to let go.

Reading

  • Matt Pocock / AI Hero

    Practical agent workflows and skills. The workflow I’m most aligned with right now.

  • Kent C. Dodds

    Testing, TypeScript, and teaching engineering well, now applied to building with AI.

  • Cursor blog

    Where AI coding tools are heading, from the people building them.

  • Anthropic engineering

    First-hand writing on agents, evals and Claude Code.

  • Theo

    Sharp, opinionated takes on the TypeScript and AI ecosystem.

  • The Pragmatic Engineer

    How engineering teams are actually adopting AI.

  • Lauren Tan

    Running agents at scale. She’s the author of pstack.