How I use LLMs

1. name of the game is 'describe expectations' (spec) and 'enforce expectations' (tests)
2. research and plan with the model (GPT-5 Pro), then implement in parallel with as many models as you can stand (GPT-5 Codex)
3. 'close the loop': let the model read logs, edit project settings, run the app, take screenshots, etc.; this is about back-pressure
4. spend more time thinking than babysitting

Tools:

• always: Codex CLI + lots of bash scripts and saved prompts
• sometimes: Xcode
• rarely: Cursive, Vibe Kanban

My philosophy:

• context engineering matters, at least for now
• see your system
• simple made easy
• design with types

← Back to home