Watch James build real apps with Claude Code, live

The audience picks an app, James builds and ships it with Claude Code on screen, and then answers every question in the chat. The full replay is here, with chapters, every tool he used and the skills you can install yourself.

Watch the latest build
Latest · 21 October 2026 · 140+ watched liveAn app in an hour with Claude CodeThe chat votes for a race-day fuelling planner for triathletes, and James takes it from PRD to a live Vercel deploy in 60 minutes, then about 40 minutes of Q&A, ending with version 2 going live.

Every live build follows the same six moves

This is what made a one-hour build possible. The times are where each move happens in the first replay.

  1. 1

    Check the idea before building it

    Before committing to the first idea, James asked Claude to check whether the sites allow scraping, whether they offer APIs and what those cost.

    Watch this step in the replay (10:03)
  2. 2

    Get interviewed instead of writing the prompt

    One line of idea, then "interview me with 20 multiple-choice questions". The answers become a PRD: what the app does, what it does not, and how to know it works.

    Watch this step in the replay (18:02)
  3. 3

    Wireframe before any app code

    Clickable HTML wireframes take seconds to change. The same change after the app is built means a server restart and a backend to rework.

    Watch this step in the replay (29:08)
  4. 4

    Let critic agents attack the plan

    Ten critics read the PRD from different angles and found 79 problems, 35 of them serious. One result: a safety screen that rules out people the plan could hurt.

    Watch this step in the replay (42:35)
  5. 5

    Clear the context, then build in parallel

    Everything goes into the PRD, the session is cleared, and a fresh Claude builds the milestones with subagents working side by side.

    Watch this step in the replay (54:37)
  6. 6

    Scan, ship, then keep testing

    A secrets scan with gitleaks and a deploy to Vercel with about a minute to spare. By the end of the session over 900 automated tests guarded every change.

    Watch this step in the replay (1:14:52)

The stack James builds with

The tools that did the work in the first build. The replay page explains how each one was used.

  • Claude Code
  • Claude Opus 5.5
  • ChatGPT / Codex
  • Next.js
  • Tailwind CSS
  • shadcn/ui
  • Vercel
  • GitHub
  • Paper
  • Magnific
  • CleanShot X
  • Perplexity
  • Wispr Flow
  • Playwright
  • gitleaks

Questions from the live chat, answered by James

Why name Next.js, Tailwind CSS and shadcn/ui in the prompt?

Giving Claude a framework saves a lot of time. Tailwind skips hand-building the CSS, and shadcn/ui hands Claude ready-made components, so menus and charts come out right first time instead of needing round after round of fixes.

Why spend 40 minutes planning in a one-hour build?

Changing a text document is quick. Once the app is built, every change can take a minute or two while the code and server catch up. James makes as many decisions as he can up front, so the build itself can run in one go.

What does a setup like this cost?

James pays £200 a month for Claude and rarely hits the limits, £100 a month for OpenAI, and between £1,000 and £2,000 a month on software overall. Fast mode is billed on top: about £20 had gone by the halfway point of this build.

Is Vercel secure enough to run a real business on?

James runs his whole business on Vercel and has been impressed in a year of using it, and some of the biggest AI companies run on it too. He also called security very important, and covers it in a whole chapter of the Agent Accelerator.

What if my app stores personal data?

Get someone to look at it before it goes live, because there are a lot of unknown unknowns. Clerk handles sign-in and a lot of the security for you, and the OWASP skill and Claude Code security review check for the classic mistakes. A marketing site that only sends emails to Kit or Mailchimp is lower risk.

Do I need ChatGPT or Codex as well as Claude?

It isn't essential. James treats Codex like an extra senior engineer: for most big pieces of code, especially anything around user accounts or security, he has it review the work at its highest effort, because a different model usually catches something. If you are paying for yourself, he thinks it is worth it.

Can I use subagents in Claude Cowork or chat?

Yes, in Claude Chat, which Cowork is merging into: ask Claude to use subagents and you get some of the same benefits. Each subagent takes a narrow task and reports back, which protects the main conversation from filling up.

Would Lovable do the same job?

Jean-Michel, the producer, answered this one in the chat: you can, but Claude Code gives you better control over the logic and lets you keep building on top of what you made.

Learn the method properly

The build method in these replays is chapter 7 of Claude Code for Builders, taken step by step on a real feature.

See Claude Code for Builders