In this session, I’ll reveal what happens when you stop using AI as autocomplete and start treating it as a long-running engineering partner. I’ll share how I used Codex across multiple worktrees to build LocalStudio.dev, a browser-native, local-first alternative to tools like Canva and Keynote, with more than 770M logged tokens and a longest uninterrupted agent run of around 20 hours.

We’ll explore the workflow behind the project: spec-driven development with Superpowers, parallel worktree-based implementation, custom Codex skills, and later end-to-end test coverage to stabilize the product. You’ll see how the project evolved from idea to working app, what broke along the way, and what I learned about using AI agents for real software engineering beyond toy demos.

Outline

- The Experiment: 770M logged tokens and a 20-hour uninterrupted agent run
- LocalStudio.dev: building a browser-native Canva/Keynote alternative
- Worktree-Based Development: running parallel Codex tasks without blocking the main project
- Spec-Driven Development: using Superpowers to guide planning and implementation
- Custom Skills: repeatable workflows for planning, testing, memory, and reviews
- E2E Stabilization: using Playwright coverage to find bugs and harden the app
- Lessons Learned: what worked, what failed, and how to run better agentic coding sessions