I stopped using Cursor as an autocomplete tool and started using it as a second engineer. Here's the shift that actually moved the needle for me as a solo founder:

Old habit: ask for a feature, get 200 lines, paste it in, pray.

What works instead:

1. Write the plan in plain English first, in a comment block, before touching the AI. If I can't describe the feature in 5 sentences, the AI can't either — it just hallucinates confidently.

2. Ask for a diff, not a rewrite. "Change only what's needed to add X" produces code you can actually review. Full rewrites hide bugs in places you're not looking.

3. Make it write the test before the implementation. Sounds backwards, but it forces the AI (and you) to define "done" before any code exists. Cuts scope creep hard.

4. Review like it's a junior's PR, every time. The moment you stop reading line-by-line is the moment silent bugs start shipping to production.

Solo devs don't have a second pair of eyes on their code by default. Treating your AI assistant as a disagreeable collaborator instead of a code vending machine is the closest substitute I've found — and it's the difference between shipping fast and shipping fast and broken.

What's your AI workflow looked like as you've scaled it up?