How I code with AI changed a lot

TL;DR

• 1

  Theo mostly abandoned plan-heavy Cursor workflows for GPT-5.5 + Codex-style apps — after building Lakebed, he says he now uses GPT-5.5 almost exclusively, rarely touches Claude, and prefers app-based harnesses like Codex or T3 Code over IDE sidebars and terminal-first flows.

• 2

  Remote coding became core, not a nice-to-have — he wants agents to keep running after he closes his laptop, and found T3 Code’s remote/browser setup dramatically more reliable than Codex desktop remote control, especially for typing latency, image paste, and stability.

• 3

  The real skill is context steering, not prompt overengineering — Theo’s key advice is to read the model’s text carefully, argue with it, correct its assumptions, and use a lightweight Agents.md plus examples and glossaries instead of giant system prompts, skills, or file-by-file instructions.

• 4

  He now prefers many short, isolated threads over long chats or parallel worktrees — on Lakebed he says he started over 100 threads in 5 days, usually one task per thread on main, because old chat history biases the model and hurts performance more than re-exploring the codebase does.

• 5

  Simple prompts often worked shockingly well when verification was in place — he shows cases where a two-sentence prompt produced a full spec in under 2 minutes and a working implementation in about 10 minutes, with Codex computer use, tests, and review tools like CodeRabbit helping confirm the result.

• 6

  His strongest claim: developers are looking at code too much and conversation too little — Theo argues that if you’re staring at files more than the back-and-forth with the model, you’re missing where agentic coding is strongest: building shared context so the model can make good decisions on its own.