11 Things I learned after using AI Agents full-time
Article Summary
Antoine van der Lee went all-in on AI agents for iOS development and jokes that 'Arie and Ingrid (AI) are doing my work while we're having lunch.' After months of daily use, his workflow has changed more than in his entire career since 2009.
Van der Lee shares 11 hard-won lessons from using AI agents like Cursor and Codex CLI full-time to build real production apps. He's building Vydio, an app for YouTube creators, while documenting his journey from proof of concept to product. These insights come from someone who learned that AI agents amplify both good habits and bad ones.
Key Takeaways
- Plan mode prevents wasted tokens and surprises by reviewing intent before execution
- Custom ChatGPT projects with codebase context drastically improve prompt quality
- Multiple focused agents beat one large prompt for reviewable, maintainable changes
- Linters, hooks, and rules act as guardrails that make agent output predictable
- Context-free PR reviews by agents catch unclear code future maintainers will struggle with
AI agents don't remove responsibility for code quality, they multiply it: every vague instruction and skipped review gets amplified at scale.
About This Article
Antoine van der Lee kept hitting the same wall: every time he talked to ChatGPT, he had to start from scratch. He'd explain his constraints, his preferences, how his architecture worked. Then the next conversation would reset and he'd do it all over again.
He trained a custom ChatGPT Project on PDFs, audience research, and Reddit threads about his codebase. This gave the model enough context to generate agent and planning prompts without him having to repeat himself.
The output got noticeably better. The model stopped filling in blanks with random defaults. Instead, it actually understood what his YouTube creator app should do and what it shouldn't, so the initial prompts came out shaped the right way from the start.
